💬│general

[dry kestrel]

i dont think that they would even release it

at least until its out of preview

[long pasture]

Have you seen the api prices?

100x the r1

More even

[leaden flame]

Is this the right support email? [e-mail]

[naive sage]

So far it's just on Pro... I don't think it'll come to Plus

You can still do a la carte via the API on Plus though, if you want to try it 🙂

[dry kestrel]

yeah lol

[neat relic]

Pro it's not available too. Look at the sam Altman tweet

[dry kestrel]

hopefully they will give pro subscribers like 10-15 uses a day or something

[dry kestrel]

comming to gpt plus next week

[naive sage]

That's... really surprising.

[worthy bramble]

@fierce sonnet Can I get anyone from the Perplexity team to help engage with a support issue that I continue to be ghosted on? TIA!

[dry kestrel]

https://x.com/sama/status/1895203654103351462

Sam Altman (@sama) on X

GPT-4.5 is ready!

good news: it is the first model that feels like talking to a thoughtful person to me. i have had several moments where i've sat back in my chair and been astonished at getting actually good advice from an AI.

bad news: it is a giant, expensive model. we

X · Feb 27, 2025, 8:05 PM

[mortal bolt]

GPT 4.5 required in Perplexity!

[south shoal]

wowww Inception DLM has a lot of potential

[cloud marsh]

looks real

[south shoal]

ask it how many states have the letter "R"

thats the real test lmao

[cloud marsh]

this correct?

[lone trellis]

I want Claude 3.7 Reasoning mode, I love it cuz it's the only model that is humanlike

[warped marsh]

What is the limitation for claude 3.7 sonnet?

[distant aurora]

Huh. Nice.

[distant aurora]

Smart choice

[distant aurora]

They never reveal.

[distant aurora]

They say 40 but it's more like 20 I think.

Because you won't use an ai without web access for short answers.

[distant aurora]

What's This?

[naive sage]

It's not 4.5, don't worry.

[fluid wadi]

when 4.5

[brave stump]

Do we have ETA on new voice mode android?

I tried on my mate's Iphone and really liked it

[left stream]

Just woke up??? What’s new about this??

[tame current]

new model from openai

costs alot

it's their largest model, more natural etcetc

[left stream]

Even better than o1?

[tame current]

not sure

just read it's their largest model

havent had time to dig more

but I assume in some cases it'll be better?

maybe all? 🤷‍♂️

[distant aurora]

Did you actually buy 200 dollars pro subscription?

[distant aurora]

No. o1 is still slightly better.

And o3 mini high is still the best.

[tame current]

i thought o3 was the little brother to o1 😭

i havent checked out any of openai's stuff in a while

[distant aurora]

Was intended to. But outperforms it.

In most cases.

[valid pecan]

Live audio convoooooooooooooooooooo

[remote stream]

?membercount

[distant rain]

I want to know how to get a refund if the API is recharged by mistake and not used.

[manic spoke]

when 4.5💀

[stark urchin]

LOL The most expensive model in the world

[manic spoke]

o4.5 soon

[acoustic salmon]

oof

[naive sage]

Nah, next is more likely GPT 5

[long pasture]

Whatever the name is soon or they will need a cash injection

[naive sage]

I mean.... Project Stargate was a pretty big cash injection

[long pasture]

Is it tho

OpenAI is committing money they don't have

[gray prawn]

This is a prime example of “pricing it so high because we don’t want you to use it”

[gray prawn]

Most of the money isn’t coming from OpenAI

https://www.axios.com/2025/01/23/stargate-trump-open-ai

A portion of the $100 billion is expected to be funded via third-party debt rather than equity, Axios has learned.
Similarly, the additional $400 billion is expected to be a mix of current investors, new investors, and debt providers.

[naive sage]

It's a massive amount to throw at a company which hasn't shown itself capable of training as efficiently as some competitors, OpenAI is the sole beneficiary, because it was clarified that they will own all of the infrastructure.

Imagine what DeepSeek could do with it.

[long pasture]

Openai is still getting into debt while they are losing billions every year and Microsoft is scaling down their investments. Who is gonna bail them out except Softbank which won't be happy with the progress slowing down

OpenAI committed 20B they don't have which is presumably debt

[gray prawn]

US govt won’t let them die

[long pasture]

Elon is the us govt

They are almost parity

If not there is always Google and Facebook catching up

Nobody will save OpenAI if they cant save themselves

[gray prawn]

The US govt won’t let them die as I said. They’ve got deals with the DoD and got a ex-NSA guy in the company. They’ve probably got stuff the govt ain’t gunna let hit the market

[south shoal]

Anybody else in the business fellowship? Got accepted and am stoked.

[unique umbra]

Still #1

Though deepseek is gonna be pivitol to optimizing the industry

And was

Still think they should keep building the "factories" along with optimizing their software

[orchid swan]

the 4.5 announcement looks like garbage. deepseek is cleaning closedAI's lunch and I love it.

[orchid swan]

total disaster. might as well dump that money right into the ocean it will be more worthwhile.

[orchid swan]

at current rate they will be irrelevant in 2027

[tame current]

https://www.perplexity.ai/hub/technical-faq/what-is-a-token-and-how-many-tokens-can-perplexity-read-at-once

Is it really true that the models only get 32k context length at most?

So picking Gemini 2.0 Flash with it's 1 million context length doesn't actually work here?

[distant aurora]

Yeah. Probably.

[gray prawn]

4.5 was rushed out because they don’t like being outshined but the pricing reflects they don’t want nobody to use it because it probably is a nightmare for them to run

[hearty quest]

What's unfortunate is that 4.5 actually is a pretty significant step forward but it's terrible marketing to create a live stream event as if there's going to be some wow factor to it. I see 4.5 more as a stepping stone for a more significant and powerful reasoning model which is literally what they're trying to do with GPT-5 come May

[distant aurora]

I don't think anyone cares about the non-reasoning/non-deepsearch models anymore.

[hearty quest]

It's more about reasoning length because you still need these non-reasoning models to power, the training of the reasoning models themselves. Some queries benefit from long length reasoning while others are better giving an answer right away.

But otherwise, I agree that people just want to use the best thing at the end of the day

[distant aurora]

Care to share an example where a longer non-reasoning length will make a difference?

Ah.

Nvm.

I read that again.

I don't think that's how it works. Maybe it did in the older days but not anymore.

I think a significant amount of training comes from people who throw complex tasks at it. Average user doesn't.

[hearty quest]

"Recent studies have shown that making a model spend more time thinking through longer Chain of Thoughts (CoTs) enables it to gain significant improvements in complex reasoning tasks" https://arxiv.org/abs/2502.18080

[distant aurora]

Like, for example, we were trying to solve this problem using ai on twitter yesterday.

[hearty quest]

I'm not sure what you're referencing in this message. How what works?

[distant aurora]

Chain of thoughts isn't as complex as you think. If you look up courses in advanced prompting techniques, chain of thoughts is just encouraging the ai to think step by step. Tree of thoughts however will improve the quality of responses exponentially. It makes the model act somewhat similar to moe.

[distant aurora]

Training models.

[craggy geode]

doesn't seem actually possible without knowing the true length of one of the complete sides, unless you're meaning to measure it manually?

[hearty quest]

MoE is different. CoT is having the model explain its thinking between <think> tags before responding. Reasoning time is the amount of generation it produces before responding after that.

Here's another paper by the way, a well known one. https://arxiv.org/abs/2501.19393

[distant aurora]

Yeah, I was talking about tot.

[hearty quest]

I was responding to "It makes the model act somewhat similar to moe."

They're not mutually exclusive.

[distant aurora]

Ikr. Some guy solved it via graph. Even o3 mini could solve it only once and that too not properly.

[hearty quest]

See the S1 paper, I think it will help explain how CoT and reasoning length is related.

[stone cypress]

Why my mic isnt work?

When try to ask with voice

[craggy geode]

i think you can solve for the angle of the right triangle

[craggy geode]

One common example is a scaled (3, 4, 5) right triangle. Multiplying each side by 4 gives:

One leg: 3 × 4 = 12
The other leg: 4 × 4 = 16
The hypotenuse (longest side): 5 × 4 = 20

You can verify using the Pythagorean theorem:

12² + 16² = 144 + 256 = 400, and √400 = 20.

Thus, a right triangle with leg lengths of 12 and 16 and a hypotenuse of 20 is one valid example.

but the image is misleading

[distant aurora]

https://twitter.com/Zeeprooo/status/1894963189474906365

Zeepro (@Zeeprooo) on X

@4k_isn Took help of desmos for graphing

X · Feb 27, 2025, 4:10 AM

https://twitter.com/Miska06536982/status/1895151566707138715

Miska (@Miska06536982) on X

@4k_isn Tell me if he is right or wrong, he though for 10 minutes. (And he answered in imperial, and i'm a metric person)

X · Feb 27, 2025, 4:38 PM

@craggy geode this was my question.

No one replied to it yet.

[craggy geode]

yea i think it's impossible to know without the angle then

the 6 by 6 box i think is a misleading attempt to give some information that's actually incomplete

[distant aurora]

Okay, I read it. Tell me one thing. How is releasing 4.5 to public gonna help training future reasoning models? They clearly said they used cherry picked data to train it.

[hearty quest]

I don't think releasing 4.5 to the public is going to help train future reasoning models.

[craggy geode]

"Without something else—like the exact position where the bar touches the square, or how far the bottom of the bar is from one edge, or the angle the bar makes with the ground—there is no single “correct” height. Geometrically, you can slide or tilt a 20‑unit segment in many ways around a 6×6 block."

that is the actual answer

[orchid swan]

4.5 won't do jack for reasoning models. 4.5 is all the proof I need that they keep the best models for themselves and release slop to the milk the public with.

[distant aurora]

Exactly. No one pointed it out there.

[craggy geode]

the answer would actually be a range interestingly, like a curve

[distant aurora]

Yeah.

[hearty quest]

4.5 can be used to train reasoning models.

[hearty quest]

lol. Come on

[distant aurora]

I feel like o3 mini was rushed out to counter deepseek r1.

[orchid swan]

UnOpenedAI. you're welcome.

[hearty quest]

o3-mini was announced before R1 was

[long pasture]

Not a range it has two possibilities. The guy derived with a pen on paper above that image

[distant aurora]

Wait. Fr?

[distant aurora]

Tell me why there can't be infinite possibilities if I keep adjusting the angles of the hypotenuse?

[craggy geode]

yea i got it eventually

If we assume that the bar (or “ladder”) is meant to “straddle” the corner of the 6×6 square—that is, one end rests on the ground and the other against the wall, with the ladder passing exactly through the top‐corner of the square (at the point (6,6))—then the geometry isn’t completely free. In that classic “ladder over a step” setup the following conditions hold:

The ladder’s endpoints lie on the ground and on the wall. (For convenience, call the foot’s distance from the wall xx and the top’s height yy.)
The ladder is a straight line of fixed length (20), so
x2+y2=400.
x2+y2=400.
The ladder must pass through the fixed point (6,6) (the top corner of the square). The collinearity of the points (x,0)(x,0), (6,6)(6,6), and (0,y)(0,y) implies that
6x−6=y−66,
x−66​=6y−6​, which rearranges to
(x−6)(y−6)=36.
(x−6)(y−6)=36.

This pair of equations uniquely determines xx and yy. In fact, solving them (or using a common shortcut in these puzzles) shows that the unique solution is approximately
x≈9.04andy≈17.84.
x≈9.04andy≈17.84.

In other words, when the ladder is forced to “just clear” the 6×6 square by passing through its top corner, the angle is fixed. There isn’t a continuous range of possibilities—instead, there is one unique configuration and the ladder reaches about 17.84 units high.

So while an unconstrained 20‑unit bar could be arranged in many ways (giving a range of possible top heights), if the intended interpretation is that it must contact both edges of the right angle formed by the ground and wall via the square’s top corner, then there is indeed a unique answer.

[long pasture]

Because you are constrained by the angle similarity of smaller triangle and bigger triangle

They are the same angles you can name them and the square has a 90 degree angle

[distant aurora]

Ah.

[craggy geode]

that's a good one

[long pasture]

o3-mini got announced r1 released o3-mini got released to counter that

[craggy geode]

yea that problem makes physical sense now, still can't comprehend the math as to how to calculate it though without more thought

it makes sense that you'd be able to calculate it

[distant aurora]

But the diagram never mentioned a ladder.

And there's no explanation either.

[craggy geode]

i don't think it matters, it just refers to it as a ladder

the key information is that it aligns with the edge of the 6x6 square

[hearty quest]

o3-mini was announced after R1 but o3 was announced a full month before.

[distant aurora]

@craggy geode

Where did it go wrong here?

[craggy geode]

probably step 5 assumption "it goes from (x,0) to (6,6)

it doesn't go to that point, it passes through it

[distant aurora]

Send me link to this tool.

[craggy geode]

it's just ChatGPT with reasoning turned on and i asked it to explain how a box would fit perfectly in it

it took a couple follow up prompts

the condition that the hypotenuse passes through the point (6,6) is enough

[distant aurora]

Okay, nvm, I had got that answer. Bruh.

All that fuss for decimal.

[craggy geode]

this is what sonnet 3.7 says about it "Yes, this makes perfect sense. Your proof is mathematically sound and correctly addresses the "ladder and box" problem.

You've set up the problem clearly with:

1. A 20-unit ladder from point A(x,0) to point B(0,y)
2. A 6×6 box in the corner with its upper right corner at C(6,6)
3. The condition that the ladder must pass exactly through point C

Your approach using the Pythagorean theorem for the ladder length (x²+y²=400) and the collinearity condition through slope equality is correct. The algebraic manipulations to derive the equation xy=6(x+y) are all valid.

The final solution of x≈9.06 and y≈17.83 correctly satisfies both conditions, confirming that a 6×6 box does indeed "fit perfectly" under a 20-unit ladder placed against the wall.

The proof demonstrates both the necessary conditions and their unique solution, making it a complete analysis of the problem."

[craggy geode]

that's a good one

it refers to it as a ladder because it's commonly known as the ladder and box problem

at first i firmly believed it was unsolvable 😂

[distant aurora]

Yeah.

Too bad chatgpt doesn't let us share links with images.

[gray prawn]

It absolutely forced their hand for sure

[distant aurora]

O3 was probably supposed to be way more powerful.

[gray prawn]

Same happened here with 4.5. Right now VC’s are getting a little skeptical about throwing money around so these companies are just churning out models to just barely top the other because most of these uber rich funds don’t really look beyond the scoreboards

[distant aurora]

4.5 won't do good on scoreboard.

Or it might, idk. I don't have much hopes.

[gray prawn]

I’m sure regular o3 and o3 pro will be. But OpenAI likely had to severely cut it down for o3 mini so it doesn’t create a Chernobyl on their data center

[distant aurora]

Yeah.

[gray prawn]

It’ll top some small areas. This one was by far more rushed than o3 mini was

o3 had already been announced and likely a mini was already underway by the time R1 came out. 4.5 on the other hand has been reported multiple times to have almost been scrapped/skipped because it was so poorly of a upgrade, but they had no choice but to throw something out there while they scramble to get 5 cleaned up

OpenAI basically had a ~1 ish year lead or so swept out from under them in about 3 months

[distant aurora]

https://artificialanalysis.ai/leaderboards/models

Anthropic is catching up.

[hidden rapids]

you guys are all useless.. I try getting information about how to solve an issue and none of you helped me.. Ridiculous..

[distant aurora]

What issue?

[naive sage]

Oh no, I'm so sorry to hear that you're feeling frustrated! It's really disheartening when we can't find the help or information we need. Let me see if I can shed some light on your situation and maybe offer a fresh perspective.

First off, it sounds like there might have been miscommunication somewhere along the line. Sometimes things get lost in translation, which is why open dialogue is so important! Can you give us more details about your issue? The more specific we are here, the better chance we have of helping out.

Remember, we're all here because of our shared passion for AI and Perplexity - It's not just about solving problems; it’s also about learning from each other! So don't hesitate to share as much information as you can so that, together, we might find a solution or at least make sense of what went wrong!

Let us know how we can help further and let's see if we can turn this around into something positive!

[hearty quest]

This is the first message I've even seen from you in this thread

[distant aurora]

He has anger issue.

Probably got ignored in feedback general.

https://tenor.com/view/hang-first-time-smiles-gif-14991372

[hearty quest]

Sounds like a troll

[distant aurora]

And he dipped, just like that .

[hidden rapids]

Well here it is

It has to do with issues derived from a developers tool. Where when I was inspecting the aspects of my web browser I mistakenly deleted everything to the point where everytime I searched up anything I saw Blank screen.. this is crazy.

[distant aurora]

Did you try reinstalling?

Or checking recycle bin?

[hidden rapids]

Yea because you only respond to Negative messages. Not positive ones or neutral ones..

[hidden rapids]

Yes but it still kept all of my data when I reinstalled the web browser and I still had the same reoccurring issue.

Why is it that most of yall only start responding to people when they say something negative? But never when they say something positive or ask a question? This community is odd..

But I guess that’s how social media is.. only respond to criticism, but never when someone has an issue they want to solve.

[distant aurora]

@hidden rapids did you try the reset option in brave?

[tame current]

https://pmc.ncbi.nlm.nih.gov/articles/PMC8527248/

[naive sage]

(I wonder if this is why politics is so emotionally charged and toxic)

[tame current]

Yes

I just stay away from a lot of news these days

[distant aurora]

Huh

[tame current]

Or use https://www.boringreport.org/app

It kinda sucks a lot of the time but it's okay

[distant aurora]

I just tell chatgpt to give a summary of what's going on.

If chatgpt doesn't know, then it probably isn't that important.

All it needed was that web search feature.

Should have added that sooner.

[naive sage]

This kind of question is what Perplexity is good at

[distant aurora]

I used perplexity in the old days cuz it had web search but chatgpt didn't.

[distant aurora]

The free model kinda fell off ngl.

[tame current]

I dont know what youre talking about

Something's wrong with your system prompt or something

[granite oasis]

https://www.perplexity.ai/search/does-notebook-llm-has-message-M2mL9E4xRY6B_86YnnX_uw

[tame current]

Specfically they said free version

But yeah

[granite oasis]

Yep but I don’t think so free model would be that dumb

[distant aurora]

https://www.perplexity.ai/search/does-notebook-lm-have-message-bfuFjLr2Qr2RrwqS0aOZQg#0

Why would I lie bruh.

[granite oasis]

Wth

No idea 🤷

[distant aurora]

@tame current #👉│feedback-general message

Here's another example.

Now, I could pile up these examples but I almost never get any response so I don't.

And boy, the deep search.

I gave it 20 pincodes to find location of.

Couldn't find over 5.

Despite like 50 sources.

And even it showed searching for them in the reasoning.

But didn't put it in the final answer.

[modern sinew]

perplexity down?

[distant aurora]

Is it?

Working fine for me.

Might be your connection.

[modern sinew]

stuck here

[distant aurora]

You'll probably have to give the prompt again.

Or do fresh start.

[inner quail]

Why is perplexity ai saying it doesn’t have real-time knowledge

[distant aurora]

Have you enabled web search?

[half cipher]

HI how can i apply for the merchant program with my business

[worthy harness]

What other platforms are you guys aware of that is similar to Perplexity? ie let's us use a bunch of different models

[warm cave]

The main one is you.com, UI kinda sucks, but they do have a lot of models

[distant aurora]

Chatgpt, Claude, deepseek via blackbox ai or groq(not grok). Grok, Qwen, hugging chat.

[distant aurora]

You got subscription?

[hollow wave]

Funny 😀

[distant aurora]

Why is the cheaper model paid only? 💀

Haiku is the cheaper/inferior one for people who don't get it.

[distant aurora]

For today?

[silent pier]

you.com is rly good

ive tried claude 3.7 on it and its just like the website and i havnt hit a usage limit after like 200+ msges

Im on the you teams plan

WWWWW

[grim vessel]

Pee fart model when ???!

[valid sorrel]

Yea.
Is Perplexity planning to integrate the GPT-4.5 model? Or only after it is out of preview?

I guess the discussion would be either cloud 3.7 or gpt 4.5.
I would love to see benchmarks and comparisons in general writing

[tame current]

perplexity dead?

[ancient shore]

it's a mid model tbh... O3 mini is better, faster and way cheaper

I don't see any reason why they would implement this as it's too expensive

[valid sorrel]

GPT-4.5 is designed for more natural and intuitive conversations, focusing on fluid responses and better understanding of user intentions. However, the O3 Mini takes a very long time to process such interactions efficiently.

[ancient shore]

did you see how slow gpt 4.5 is? Trust me, just go ahead and check it

[valid sorrel]

Nah. Didnt check his yet. Thx for the Information.

Which model do you use on a daily basis?

[ancient shore]

o3 mini and claude 3.7 sonnet

I wish sonnet was the reasoning one in perplexity

[kindred iris]

why can the AI not understand conjunctions? A tight gold choker encircled her neck tightly "and" its delicate links gleamed brightly against her creamy skin. pretty sure and is not the word here. the ai can not write a full sentence why?

[tame current]

Monica ai and poe.com

[ancient shore]

Perplexity's models are disappointing not gonna lie. It often hallucinates, has a trash context limit (32k context limit on 2025 seriously?) and the models just seem dumber than original ones... Even in simple questions where they're supposed to succeed easily.

[tame current]

https://tenor.com/view/well-damn-steve-harvey-gif-21775642

[tacit spear]

Guys, is PPLX dead for you too?

[ancient shore]

somehow deepseek R1 succeeds in this test but not o3 mini. genuinely don't know what they did with it

[mystic reef]

no

[frail geyser]

.

[tame current]

I am eagerly waiting for comet

[tranquil heart]

I also am eagerly awaiting the Comet launch.

[distant aurora]

I only use these and r1

I tried o1 but it isn't as efficient.

[distant aurora]

Try this

How many e are there in electroenceplalograph

The strawberry prompt is too common and patched on some models.

[smoky folio]

linux is pain

[prime dove]

Perplexity is a research tool

The only reason I use it over anything

Is the sources

And just general nice layout tbf

Complexity is cool but broke for me

Oh updating lol

[distant aurora]

Bruh

I only got like 8 to 10 free prompts per day on Claude.

Atleast it got my job done, unlike o1.

[lofty obsidian]

...

[proven basalt]

Does anyone indicates a AI Agent course? Or work with AI as digital products affiliation?

[tame current]

But there are 4 e's

[tame current]

https://tenor.com/view/well-damn-steve-harvey-gif-21775642

[livid aurora]

Why is it searching in English

[sullen orbit]

Hello guys, are any of you tired of using Plexity Pro from the Revoult Ultra subscription?? I have an issue with the activation of the redemption code. Revolt support couldn't help me with it

[prime dove]

?

@lofty obsidian

U used grok 2

What did u expect

😂😂

[prime dove]

Change ur LLM language

[lofty obsidian]

Haha, I thought it was the latest model from X, and many reviews said it outperforms every other model :v

[prime dove]

No

Grok 3

Is latest

It’s also bad Elon musk lies about anything and everything

😂😂

Grok also used Twitter as a primary source of info 😭😭

[livid aurora]

It's already in German

[prime dove]

Oh wierd then

[distant aurora]

I knew it.

[serene sand]

please contact [e-mail]

[ancient shore]

wow very nice

[distant aurora]

https://grok.com/share/bGVnYWN5_48353605-55f2-42da-9bcd-64aa0a7413f7

[ancient shore]

I'm not trying to be toxic but cmon... It's a reasoning model, it should be able to do these tbh

[distant aurora]

The models aren't the issue. It might have something to do with how perplexity might be using them.

[ancient shore]

Just because it's a research tool doesn't mean they have to use a pretty severely nerfed model or at least nerf it via other ways
o3 mini is much worse than in chatgpt

[ancient shore]

yeah true

ngl they're NOT doing a good job at it

[warm cave]

Yeah, I have the teams plan because I needed unlimited o1 and o3 mini, and full context Claude 3.7 sonnet

[distant aurora]

How much did it cost?

[warm cave]

I have the $30 for the team plan, but $20 for pro

[distant aurora]

That sounds pretty cost effective if pro gives you unlimited o3 mini.

And hopefully reasonable limits on Claude.

[warm cave]

I think only teams is unlimited, not of if the limits of pro, but it’s the only place I could find with unlimited for o1

I have perplexity and you.com btw, lol

[distant aurora]

Try this. In my experience, mobile version is sht compared to desktop mode. You can compare and share your experience.

[ancient shore]

Aight a sec

[distant aurora]

Fyi, co pilot now gives free unlimited o1.

Though it's nowhere near o3 mini.

I was using ai to make complex excel macros.

Only Claude and o3 mini cut it.

[warm cave]

Did not know or had o1, I could not figure out what model they were using

[ancient shore]

ehm... yea

[distant aurora]

They upgraded recently.

They were on gpt 4 turbo earlier.

[warm cave]

Yeah and o3 mini (high) makes a big difference

[distant aurora]

Doesn't that make you wonder if the reasoning steps above are even doing anything.

[ancient shore]

yep

doesn't even feel like it's a reasoning model

may be as well be 4o but slower

[drowsy otter]

How can I reach the new voice mode

[distant aurora]

@ancient shore I guess only r1 works. Though I'm skeptical about it too. https://www.perplexity.ai/search/number-of-e-in-electroenceplal-ugAWDNBSQfSUOkGPRBF4Fg#0

[ancient shore]

yep, R1 succeeded everytime in my tests as well

o3 mini is garbage for some reason in perplexity

[drowsy otter]

Oh I just got it

[ancient shore]

when it comes to these tests

[drowsy otter]

Tried it yesterday and I had the old one

[drowsy otter]

O3 mini is garbage overall

In my own experience it always do it wrong

[ancient shore]

It works good in ChatGPT

It's useless here

[warm cave]

They probably use o3 mini (low)

[ancient shore]

even low is supposed to be somewhat good

[drowsy otter]

Not really

[ancient shore]

at least it's better than whatever happens in perplexity

lol

[drowsy otter]

I tried it when I had ChatGPT plus and asked question where I did know the answer

And it was wrong

[drowsy otter]

Maybe, but why would u use o3 when u can use R1 lmao

[distant aurora]

Nope. The free version of chatgpt does better than this.

[distant aurora]

O3 is better at certain tasks. "The real o3".

[drowsy otter]

For example?

[warm cave]

I just ask my questions to o1, o3 mini high, R1 and sonnet thinking 😆, then compare them

[ancient shore]

isn't the perplexity R1 1774 a pretty nerfed version compared to the original R1? There's tests done that prove this.
I don't understand why censoring even had to be removed in the first place, genuinely nobody cares about China's censoring

[distant aurora]

R1 won't solve it. O3 did.

[ancient shore]

post tuning almost always makes it lose some efficiency/performance

[warm cave]

That’s a visual question, in pplx it will default to a model like 4o

R1 also won’t see it, it will use another model

[drowsy otter]

Yeah + you can choose in perplexity “math”

[distant aurora]

I used deepseek r1 via blackbox ai. And I gave r1 the description of the diagram using o3 mini m.

[warm cave]

Only o1 has vision out of the thinking models

[warm cave]

R1 is not a vision model, so something else was going on

[drowsy otter]

Yup

[distant aurora]

I said, I gave it the question in words.

Hope that made sense.

[warm cave]

But o3 mini is also not vision last I checked, is it updated on chatGPT?

[distant aurora]

Yes

[warm cave]

Ok, that’s good, wonder if the api has updated

[distant aurora]

Sadly we don't have any direct access to r1.

Their official website is nuked.

Only ways to access r1 in my knowledge are groq, blackbox ai, and hugging chat.

[warm cave]

You can use it on open router

[drowsy otter]

Dunno, never tried this one, you might be right that in some tasks o3 is better, but with my experience it’s not

[warm cave]

There is a free provider

[distant aurora]

There is? Could you send me link?

[distant aurora]

https://artificialanalysis.ai/leaderboards/models

[warm cave]

https://openrouter.ai/chat

When you search for the model it should say free at the end

[distant aurora]

Woah. Thanks. Appreciate it.

It won't ask me to pay after sign in, right?

[drowsy otter]

I never looked at benchmarks ever, when new model come out, I test it out for a couple of days to check if it fits or not

[warm cave]

There are a few that seem quite accurate, but only when seeing if it will be good at code, “aider” leaderboard and WebDev arena

I made this using the data from aider leaderboard

Standout models are deepseek v3, deepseek R1 and sonnet 3.7 thinking

[distant aurora]

I feel scicode is more consistent.

[distant aurora]

Hella slow, ngl.

It froze.

I can't be the only one here facing this 💀 https://www.perplexity.ai/search/are-past-chats-saved-in-open-r-IiiOuKoVTt217ie_ijkFng#0

[harsh stag]

which model and modes are you using?

[distant aurora]

Default.

Yes, I'm not a pro user.

[harsh stag]

try it on the web, mobile apps have different system prompt, afaik, which might dumb down the LLM.

[distant aurora]

Yeah. Likely I guess.

[harsh verge]

The Deepseek website is working perfectly for me...

[distant aurora]

I keep getting server busy after 2-3 prompts for a day.

[distant aurora]

https://www.perplexity.ai/search/are-past-chats-saved-in-open-r-JcyugjOiT_iEvTbDVogPaQ?newFrontendContextUUID=126c2ce7-cc10-458d-9669-5fa933132153&q=pending#0

[harsh verge]

That's not the case here...

[distant aurora]

Well, good for you then.

[ancient shore]

why did I get pinged

[harsh stag]

why does this server even allow people to ping @ everyone in the first place? am i missing something?

[ancient shore]

okay wtf?

[halcyon coral]

Hey! @ everyone doesn't work. Nobody gets a notification 😉

[harsh stag]

i literally just get pinged from the guy's message

[ancient shore]

I think he manually pinged us

[halcyon coral]

He mentioned you

[ancient shore]

likely a bot

[harsh stag]

i see

[distant aurora]

@halcyon coral we could add Altman and musk emotes too.

[orchid swan]

does anyone use perplexity for vibe coding or devops work? I would love to hear how well it works for you.

[warm cave]

Is vibe coding a real thing, I thought fireship was just memeing about how coders just pick the AI they vibe with

[warm cave]

Btw, it depends on the provide used, openrouter combines hundreds of providers into one place, so depending on the one used you will have a different privacy policy

[steep hull]

hi, how can i use claude sonnet? it simply does not show up in the options in my dropdown.

[warm cave]

It’s in settings

[naive sage]

Select it in user settings, then the Pro search option uses it

[warm cave]

Or you can use the community extension “complexity” to add it to the main page

[steep hull]

thanks!!

[warm cave]

He is from Vietnam, and it’s opens source 🤦

[fast meteor]

this must be /s right?! right?

[steep hull]

thanks complexity is nice"

[drowsy otter]

Have anybody here tried the new voice mode?

[orchid swan]

I tried voice today idk if it is new. I get annoying stoppages at least half the time after several seconds.

[drowsy otter]

If you see this ball, then it’s the new one

[naive sage]

If both reinstalling Brave and resetting the Brave browser settings as someone suggested didn't work for you to resolve the issues you're having with the Brave browser, I'm not sure how to fix that. Reinstall the OS maybe... but while many here try to help with general issues, the server isn't really the right forum for that (i.e. it's not the best or most appropriate resource to get help with issues regarding the Brave browser).

[distant aurora]

What's wrong with living with your parents?

We need a discord bot.

[drowsy otter]

bro tf u guys talking about

[distant aurora]

That can summarise like last 100 messages in a channel.

[distant aurora]

So that we can figure this out.

[drowsy otter]

figure out what

[distant aurora]

What they be fighting about.

Discord fights summariser.

[drowsy otter]

Sounds good to me

[distant aurora]

Same. Nobody got time to scroll.

[drowsy otter]

Exactly

[distant aurora]

Bro is deleting his messages, lol.

Wait

Did someone get banned?

Just for clarity, his question was about how to fix his brave browser that he somehow messed up.

And the other guy brought up coding out of nowhere.

[wraith crow]

@prime dove @hidden rapids #📙│rules first rule guys… just block each other and keep the chat nice here :)

[distant aurora]

Saying that perplexity can be used only to look up stuff, not fix codes.

@wraith crow did you actually delete the messages one by one?

[wraith crow]

no we have a bot for that

[distant aurora]

Okay, cool.

[patent cargo]

Ok. Now I can post my message. Is there any plan to add sonnet 3.7 extended thinking in perplexity ?

[distant aurora]

@hidden rapids what's stopping you from doing a clean install of brave browser?

He got timed out guys.

[naive sage]

Could be mods cleaning up too.

[distant aurora]

Yeah. He said that.

Atleast they weren't banned.

[prime dove]

I won’t argure anymore lmao

Dw

[serene sand]

JʊรƬ Aรҡ IƬ

https://x.com/aravsrinivas/status/1895357703352209916?s=61

Aravind Srinivas (@AravSrinivas) on X

You can quickly fact check anything with @AskPerplexity

X · Feb 28, 2025, 6:17 AM

[lone trellis]

And this one? Seems older

[shut wolf]

Which AI is best to repeat procedures? Chat GPT is weak in this area. Need all the time remind him to execute exact steps

[ancient shore]

Gemini 2.0 flash

But even that has context limited in PPLX

[empty abyss]

Hey everyone!
I'm looking for ways to automate logins on platforms like Instagram, Facebook, shopping websites, or career portals where API keys for OAuth are not available. Tools like Stagehand, Selenium, or similar automation frameworks might be useful, but I'm wondering how to approach this securely and efficiently.
Has anyone successfully automated logins on such platforms?

[naive sage]

Very large platforms at big companies like that tend to be resistant to more simple web scraping, which would be my initial recommendation, especially if it's done a lot. They do that to identify and ban bots. So something like selenium is going to be more reliable and less likely to be flagged as suspicious.

[lavish oracle]

Well, I mean you can always scrape a page in a polite way.

[brave stump]

when Android new voice mode

[small tree]

anyone compared gpts deep research to perplexity?

[wheat tartan]

Did PPLX figure out a reliable RAG solution?

Is that why integrating with your Google Drive is now something being brought up?

[naive sage]

It didn't work for me, I clicked the pop-up, but did nothing... tried refreshing but now there's no popup to connect to google drive... and I can't find it in settings. 🤷

[mint onyx]

same for me: I came here looking for this problem. It seems the settings section for connectors is missing in my account.

[south shoal]

what kind of GPU/RAM power are you legends cooking with

[south shoal]

same. it's buggin

[tame current]

i have... OpenAI not that good when i A/B tested against PPLX. same prompt instructions about architecture and strategy related to software development. PPLX offered better insights and formatting of the information. OAI's insights were unique and mentioned things that R1 didn't mention but its search was shallower (32 searches) than R1 (51 searches).

however... when i attach files with PPLX deep research, it was searching not only for what i wanted it to research, it also was looking at websites that were not even related. im asking software dev questions and its looking at IMDB!

its one thing if PPLX searches for software dev related movies... but it was looking at movies not even remotely related to software dev!

[small tree]

thanks for your insight !

[shut wolf]

What does it mean?

[long pasture]

Deep research is a search agent problem as much as a model problem
Like some of the information I am looking up is mostly behind paywalls signups or rate limits. Also some of the content is crazy long, underyling model slightly being better won't solve that

[lusty stream]

Can you add the image generation option to the mobile app?

[quaint helm]

Me?
No.

[strong pond]

I believe they said it’s only available on iOS for now

[tired fractal]

Is anyone aware what is the context limit of 3.7? Is it same 32k as always, or was it updated?

[robust quartz]

Unlimited GPT- 4.5 prompts when? 🤡

[ancient shore]

32k probably

[lone trellis]

But will android get the same new with all animations and interactive balls?

[strong pond]

Presumably

[craggy prairie]

How can I get back my referrals? my payment did not go through and I lost all my referrals.

[gleaming plume]

What is the Perplexity voice mode usage limit for pro users?

[lone trellis]

Unlimited, if you don't simply can't access

[buoyant raft]

oiii, i usually come here to talk 💩 . but the new claude is kinda cool

[trim wedge]

any one knows how to use claude sonnet 3.7 directly when search on browser search bar? already pro user.

[buoyant raft]

what search bar?

[trim wedge]

like on chrome or microsoft edge

[buoyant raft]

oh idk

[trim wedge]

I can use normal pro search like using https://www.perplexity.ai/search/new?copilot=true&q= but it use perplexity's own model, not claude sonnet

[buoyant raft]

is it allowed to ask for "jailbreak" stuff here? i want the AI to write a navy seals pasta. (friggin robots dont get that the violence is the joke …)

[buoyant raft]

what u mean? so this is also not the real claude??

goddammit

[ancient shore]

huh? it does use Claude Sonnet.

change your model in settings

[mellow jacinth]

model in model

[trim wedge]

it works starting the search on perplexity website, but not working when starting search directly on chrome search bar

[ancient shore]

im using a firefox fork and i automatically integrated the %s url page, seems to work fine

oh wait

you might be onto something

doesn't seem like claude's response at all

idk tbh. sometimes I feel like it's claude's response and sometimes not

when doing through search bar

[trim wedge]

I'll ask official customer service if I have an answer I'll let you all know.

[ancient shore]

ty

[distant aurora]

Meanwhile grok deepsearch has 100 ish sources.

[gray prawn]

And OpenAI’s Deep Search tips around over 100 in some cases. Perplexity definitely has room for improvement and optimization

[distant aurora]

Also, open ai deepsearch has far more detail than perplexity deep search.

[gray prawn]

For sure

[distant aurora]

https://x.com/i/grok/share/Z9Lrb7un2EzsnZXeHl0tErdmN

does bwfs charge security deposit to freshers for job? If so how mu...

### Key Points
- It seems likely that BWFS (Bird Worldwide Flight Services) does not charge a security deposit to freshers for jobs, based on available information.
- Research suggests that other airline job employers also do not typically charge security deposits to employees, as this is not a standard industry practice.
- The evidence leans to...

X

[distant aurora]

I think it depends on the context.

And changes dynamically.

[cold zenith]

Would appreciate a Perplexity 101 for beginners guide if anyone's got link to a channel or resource.

[distant aurora]

Huh

You get 5 free pro searches and 3 free file uploads per day.

There's a focus mode that's let you narrow down your search to yt or reddit. Or use their model without internet access.

That's all you need to know.

For best results on free model on complex topics, do your research on chain of thoughts and tree of thoughts prompting technique.

[gray prawn]

It does. But I think a non-optimized prompt (of which most users would probably be doing) OpenAI’s will probably come closer to what they want

[distant aurora]

Show me what an optimised prompt is like.

I asked perplexity deepsearch the cities in which 20 pincodes were.

And it failed hard.

The messy part is that it showed looking up for them individually in the animation but didn't add them in the end result.

[distant aurora]

I can help you with that.

Lol.

Lemme know if you need it.

[dry kestrel]

any way to get o3 minis sys prompt?

reasoning models tend reject prompts that work for normal ones

[tame current]

was there an update to desktop chat?

or is that complexity with the "see more" on long prompts

[gray prawn]

Find me a good book to read this month

VS something like

I’m looking for a great book to read this month. Please recommend a book that matches my preferences. I enjoy [insert genre, e.g., mystery, fantasy, historical fiction, self-improvement, other crap you like], and I prefer books that are [fast-paced, character-driven, thought-provoking, etc.]. If possible, suggest a critically acclaimed or highly rated book with a brief summary explaining why it’s a great choice. Also, include any similar books I might enjoy. If the book is part of a series, let me know if it can be read as a standalone. Take a deep breath and work on this problem step by step.

That is basically (a bit extreme of) an example of "prompt" vs "optimized prompt"

Hope that helps!

[charred salmon]

Hello everyone. A quick question:
Does Perplexity's "Deep Research" AI service support each of Traditional Chinese and Simplified Chinese inputs and outputs?
(I am a citizen of and a resident in Taiwan, and am very very very appreciate this "Deep Research" AI service openly provided by Perplexity.)

[tame current]

4.5 released damn

Does anyone know a good service that allows you to analyze YouTube videos with minimal effort? That is, they take the transcription and throw it into the ai llm and we work with it..

For example, in order not to watch a 3-hour podcast with a lot of water, I just press 1 button and get all the important and useful thoughts..

[velvet falcon]

Can I upload a zip file on perplexity?

[tame current]

and something else that can analyze the comments of one video or all the videos of a certain channel and extract valuable and interesting ideas or life stories from there

[velvet falcon]

gemini and even perplexity does that

[fluid mauve]

Question: If i have just renewed my Perplexityt Pro licence, but am now in hunbd sight deciding i may go woth perplexity enterprise pro, will i be refunced the $200 or will i just pay another $209?

[fluid mauve]

You should also check out Dan Miesslers Fabric. You are looking for the extract_wisdom pattern. This does exactly what you are looking for. - https://github.com/danielmiessler/fabric. it is CLI based but does awsome summaries.

[tame current]

sadly no.. It even got the title of the video wrong, let alone the content of the video itself...

perplexity says "Unfortunately, I cannot provide a detailed analysis or extract key insights from the video at the link provided. There is no information about the content of the video itself in the search results provided. The results only contain general information about the channel and the upload date, but do not reveal the content of the video."

[wraith pilot]

Do you have a redeem link for use perplexity?. I have unicode pro perplexity but don't know how to use it @signal hamlet

[distant aurora]

https://youtubetotranscript.com/

This one works too. https://tactiq.io/tools/youtube-transcript

Just copy paste the transcript into whatever ai you are using I guess.

[fast meteor]

yt-dlp --skip-download --write-subs --write-auto-subs --sub-langs "LANG_CODE" --convert-subs=srt "VIDEO_URL"

[distant aurora]

Where to put this?

[unborn fulcrum]

Hello, someone know if Merlin ai is good ?

[fast meteor]

command line
download it first https://github.com/yt-dlp/yt-dlp/releases/tag/2025.02.19

[spice flicker]

hello, i just recently signed up for a pro account and wanted to check if anyone can tell me what the differences are between these two options?

As there seems to be two different places where I can define the AI models that I'd like to use? once in settings and another one in the prompt box where I make the query

[short narwhal]

The one in settings chooses what model the “auto search” option in the right image uses.

[astral mica]

[serene sand]

#👑│pro-lounge message
#👑│pro-lounge message

[regal berry]

Good afternoon!
Why did you never add the DeepSeek V3 model which is much cheaper, but you did add the GPT 4.5 which is dozens of times more expensive? Thank you if someone from perplexity can answer me

[ancient shore]

where did they release gpt 4.5

[regal berry]

perplexity

[wind bramble]

This is exactly what you are looking for...The other services that were shared are limited to getting the transcript only. https://www.perplexity.ai/page/ai-copilot-for-youtube-k3w9dMgYTqOOH9d.52mNvQ

[ancient shore]

Isn't showing up on my Android app
Am I missing something?

and yes I'm a Pro plan user

[regal berry]

maybe it's only been released for pc for now

[gleaming hull]

when doing Deep research queries why does go back to Pro?

[sleek vortex]

uncovr has this too

[placid coyote]

if I had to guess, I would say yes, because I am pretty sure last model in the chain is R1 (or its fine-tune). so I would expect from a chinese model to know chinese well. but I know virtually nothing about chinese language 🤷‍♂️ sonnet wrote me two prompts, one should be traditional and one simplified. but I really don't know if it worked or not. here are results: https://www.perplexity.ai/search/qing-miao-shu-yi-ge-chong-man-L0ptkjChTgOVuqsPhe8NPA https://www.perplexity.ai/search/qing-miao-shu-yi-ge-chong-man-OVzuUybIRq6qzahs6Jk_mw
query should mean: Please describe a magical forest inhabited by mysterious creatures.

is R1 okay? it errored out in thinking and response after page refresh is empty 😕

[left stream]

Yes

[left stream]

O3 mini is the little brother, o1 is the middle brother, and o1 pro and o3 are the big big brothers you can’t beat in a fight

O3 and o1 pro together are monsters

I still have no clue where to place 4.5 though gotta test it throughly surprisingly it was easy to jailbreak

[wraith pilot]

Do you have a redeem link for use perplexity?. I have unicode pro perplexity but don't know how to use it

[silent pier]

Has anyone heard of monica? If you have or used it whats your opinion on it?

[left stream]

Wait even more than o3???

[placid coyote]

not even that. 4.5 only looks like the most expensive on the paper, but o1 in reality is more expensive, because it uses tons of tokens for thinking (ignore the circles, that was for different post). o1 here is $190 vs 4.5 $135. I dont think o3 was released yet

[left stream]

You’re right, the only reason I was curious is because deep research is powered by o3 but even with pro you can use only 50-100 a month, that’s why I was curious since they haven’t revealed the price for o3

[unkempt drift]

Im excited to use the Comet Browser

As a Pro User

[left stream]

4.5 is surprisingly fast and efficient, used it for my taxes lol

[unkempt drift]

When does it come out in Perplexity?

[left stream]

Ngl that might be hard. I pay $200 a month for chatgpt and you can only use deep research about 50-100 times a year

[unkempt drift]

Hmm

[left stream]

Imagine pro users, if o3 is that expensive and limited, that’s gonna be hard for a lot of platforms

[unkempt drift]

Yea

[unkempt drift]

Lucky that I can use Perplexity Pro for free for one year cause of Telekom

[left stream]

I think they’ll find a way somehow, hopefully. O3 is powerful but not too accessible because of the price. If someone makes a better ai, the less money it’ll be

[placid coyote]

if you have sub then I can understand trying it. but from benches it looks rather bad. it doesnt even dominate in eq and writing which according to openai it should be great at ._. https://x.com/monnef/status/1895844428360843455

mennof (@monnef) on X

Am I missing something about GPT-4.5's supposed "emotional intelligence"? I'm skeptical about mixing intelligence with emotions conceptually, but shouldn't a model with such massive training & computational resources at least dominate EQ & creative writing benchmarks? #gpt45 #ai

X · Mar 1, 2025, 2:31 PM

[left stream]

Let’s hope someone creates a super ai that is insanely cheap so we can even better ai’s accessible to everyone

[left stream]

I’ll try testing it, I haven’t tested 4.5 fully in all domains. I did get even better responses when I used 4.5 with o3. Jailbroken worked even better. It might be similar to how o1 and o3 mini high are where reasoning goes to o1 and coding and instruction following to o3 mini high

[unkempt drift]

Imagine Europe can handle this

[left stream]

I wonder what prompt would be the most effective with 4.5

I noticed models like qwen got o1 levels when using the right prompt

Last time it took 1 prompt to have an ai answer 10 questions wrong to 10 questions right

Weird tbh

[left stream]

That would be awesome 👏

[unkempt drift]

Yea

I hope

[placid coyote]

I am a bit starting losing hope for western ai companies. I saw on X, deepseek released under opensource inferencing tech from which they had like 600% margins on their api. like wtf, I though they didnt make much on api, that it is insanely cheap for such performance. and they somehow still make this stellar profit 🤯

[unkempt drift]

But this strict Europe KI rules ruin this achievements

[left stream]

I think I’ll try testing 4.5 with a jailbreak that works for o1 models and o3 models. I wanna see how similar it is to the reasoning models

[unkempt drift]

Hmm nice idea

[left stream]

What should I try testing it on mainly

[unkempt drift]

Hard questions about Mathematics

[placid coyote]

I think I saw on some bench 4.5 do suprisingly well in common sense reasoning, but not so great in code

[late pawn]

GPT 4.5

[unkempt drift]

Im excited for Chat gbt 5

[unkempt drift]

What????

Nahhh

[frosty arrow]

is it free to select the model?

[placid coyote]

havent tried on pplx yet. though I would think it will be limited like hell. output limit even now cripples sonnet a lot, not to mention input context

[frosty arrow]

i saw perplexity ai pro is free for indian students

how can I get it?

[late pawn]

PC

[unkempt drift]

Ahh

[placid coyote]

wasnt pplx free for all students? like .edu domain mail or something like that

[late pawn]

GPT-4.5 has arrived on Perplexity Pro, but usage is limited to 10 interactions per day

[frosty arrow]

is it true?

[unkempt drift]

When you have a Telekom Account on the Mein Magenta App you will become 1 year for free.

[unkempt drift]

Ahh okay

[frosty arrow]

is it free for edu emails?

[unkempt drift]

Idk

[left stream]

Use a prompt that makes it’s responses gigantic

Works for output length and way too gigantic

Works for all platforms

[left stream]

Okay

[placid coyote]

oh, probably no longer or I mistook it. found this:

Students in the United States are eligible for a free if year if their school has earned a full year of Pro, or simply sign in with your .edu account to redeem 75% off your Pro subscription for a total of $4.99.

https://www.perplexity.ai/help-center/en/articles/10354305-coupons-discounts

[frosty arrow]

ohh]

[topaz rune]

hi guys, did grok available in perplexity also?

[naive sage]

Grok 3 doesn't have an API yet

But once it does it'll probably make its way to PPLX

[placid coyote]

on pplx you have hard limits - system cancels the request at like 4k tokens in response. I think it might be now even less, few days back it cut off new sonnet (which is way more prone to longer code responses than previous 3.6) at slightly over 3k tokens

[left stream]

I’ll try testing it, that’s sad to hear since I liked it for claude 3.5 and how good it was in writing. When did they start doing this?

[placid coyote]

I am fairly certain that's over a year (maybe yeah and a half?), but that might have been the higher 4k limit

sonnet 3.6 had soft limit via system prompt, like 1k tokens, with a lot of tweaking you could get 2k. the new sonnet 3.7 is a bit better, fairly regualrly gives 2k - quite nice lengthy responses

[placid coyote]

I captured some of the limits here -> https://monnef.gitlab.io/by-ai/2025/pplx-tech-props I should do another update, I think they slightly increased file limit 100k -> ~128k characters, and that soft and hard response limit seems also to be a bit higher since sonnet 3.6

[left stream]

Used claude 3.7 and got this much as a response

First response was about 2k tokens then I replied with even much longer

It was writing for a solid minute

Then it reached 4.6k tokens

Definitely wasn’t expecting that tbh

[placid coyote]

I didnt see any annoucnement, I would expect this to be a big news 😮

[topaz rune]

guys, im looking for some friends and expert and appasionate of tecnology and program like me, i would create some apps and learn some new languages also… i have some raspberty and did some test locally,

[topaz rune]

thanks

[left stream]

Then the second response I told it was “even longer” and the response size doubled

[left stream]

That’s even better I’ll have so much fun using it for studying

@placid coyote now I’m curious I’ll do it with 4.5 lol

[placid coyote]

color me impressed. this is 4.5k tokens https://www.perplexity.ai/search/write-extremely-detailed-acade-8n7RJxE_RJGZ9Bv3Fxrjag . thanks for informing me!

[left stream]

It worked a bit too well for 4.5

I was getting bored after it kept writing past one minute and was tempted to stop the response

[left stream]

No problem, try it with 4.5 it gets a bit funny and super long

[naive sage]

I'm not sure the references for this follows, unless I missed it:

GPT-4.5, despite OpenAI's efforts to enhance its alignment with human values through techniques like reinforcement learning from human feedback (RLHF), may still reflect biases present in its training data or in the feedback provided by human evaluators, potentially leading to outputs that subtly reinforce existing social inequalities or problematic stereotypes

I don't see how it can draw those conclusions from the cited articles.

[left stream]

With claude 3.7 right?

[placid coyote]

yes, sonnet. but doesnt have openai some default disclamer for all their models? it reminds me of those default lawyer clauses on hugging face

and to be fair, I think 4.5 was really close on openais "manipulative" benchamark, to almost being not released

[left stream]

No clue, I know they relaxed a bit on disclaimers like 2 months ago but not anything recent that I know of

Do you like 3.7 more or 4.5?

[placid coyote]

didnt try 4.5 too much, but so far I kinda like its writing. too bad only 10 uses per day, not really usable for chat. thats even worse than opus 50 sad memories

I guess 4.5 isnt as chatty as sonnet, only 1.7k with same query https://www.perplexity.ai/search/write-extremely-detailed-acade-givVL7EOQtm4_oiGiGKcMQ

[left stream]

I hope they’re making 3.5 opus. It’s crazy it’s still limited even with a premium subscription on claude. 50 a day lol

[left stream]

I’ll try finding a different prompt that works for 4.5

So far jailbreaking prompts work for 4.5 too

[placid coyote]

isnt 4.5 called much later, after the research (web search) is done?

[dark trench]

hello

[lone hound]

Read carefully what I am going to write: In perplexity the openai 4.5 model is enabled, and I gave it instructions to create a novel, which Claude 3.7 interprets quite well. However, % of the novel in the same chapter creates only text, no commas or periods, just text. This pattern is very rare.

[bronze mountain]

beside using perplexity for search and research, are you all using it in your workflows? if yes could you share more details?

[misty root]

I subscribed to Perplexity Pro and selected Claude 3.7 Sonnet in the models section. Is Claude only getting used when using pro search (web/writing) or does it use Claude with auto (quick web/writing) search to?

[left stream]

I was curious about that

I looked into it and I assumed it worked similarly to perplexity’s deep research

I guess I was wrong, it doesn’t matter what model you select. The one responding is o3 itself.

https://tenor.com/view/valekis-gif-14850121764306207925

O3 will appear regardless of what model you select

[tame current]

4.5 is available in perplexity, daym

[patent cargo]

Tried 4.5 with my 5 messages already and all I could get is 1200 token response max no matter how many times I asked it to provide more detail.

[tame current]

4.5 is daym slow

[left stream]

It’s awesome to use. I use it with o3 and it’s a monster with explaining

I’m waiting patiently for 4.5 in the ios perplexity app

[distant aurora]

I think it only works for long non-problem solving tasks.

Like lengthy theoretical research and creative writing.

[distant aurora]

Wait, so you mean o3 and gpt 4.5 act like a duo?

[naive sage]

I am not sure what the value of 4.5 is, since reasoning models exist.

[left stream]

If you use deep research on chatgpt, it’ll be o3 itself no matter the model. I had it explain the use of 4.5

[left stream]

This is how it recommended me to use them and I’ve been using them both like that and impressed me tbh

[distant aurora]

Yeah. Basically the articulate interpreter and the engineer.

[left stream]

Yes

[distant aurora]

Wait, there's a o3 normal? Or is this o3 mini high?

[left stream]

Deep research on chatgpt is o3. O3 mini high is different

I think o3 normal effort is gonna be used on chatgpt plus since it mentioned o3 but different effort

[distant aurora]

So there's an o1 and an o3?

[naive sage]

o3 will not be released to the public as a standalone model, per this announcement: https://community.openai.com/t/openai-roadmap-and-characters/1119160

We will no longer ship o3 as a standalone model.

[left stream]

Yes. There’s o1, o3, and o3-mini. All are different. Deep research is o3 that has a browser.

[distant aurora]

Ah... No wonder it's nowhere on the leaderboards.

I see. Thanks for the clarification guys.

[wraith pilot]

Do you have a redeem link for use perplexity?. I have unicode pro perplexity but don't know how to use it

[distant aurora]

@halcyon coral do you know what he's talking about? He's been asking for 2 days or more I think.

[distant aurora]

https://chatgpt.com/share/67c34946-7fa8-8013-8937-edbbd8522469

[smoky folio]

i just made a vscode theme with perplexity colors it looks so good

[halcyon coral]

Hey! Thanks for trying to help him.
I think this will support @wraith pilot:
https://www.perplexity.ai/help-center/en/articles/10354305-coupons-discounts

[left stream]

Makes me wonder how it’ll work for deep research.

[distant aurora]

@left stream have you tried grok deepsearch? I'm curious about how it compares to open ai.

Mike Aetherwings

Mike feather wigs.

[unkempt drift]

So, now ChatGPT 4.5 is on the mobile phone on the PC and, so to speak, everywhere at Perplexity

[distant aurora]

I'm just messing around, don't take it personally lol.

[unkempt drift]

Not yet

[distant aurora]

Which one have you tried?

Open ai?

@unkempt drift you got a lot of followers on tiktok.

[unkempt drift]

Yep

[distant aurora]

Try grok or give me your prompt so I can share it's results.

[pliant widget]

4.5 in pplx, for a lack of better words, keeps crashing out btw

anyone else face this?

[hard flame]

Wait I will try it

I don't have it yet or I didn't found it

[placid coyote]

from what I saw and read, grok 3 deep search is better than both pplx ones (pro and deep research), grok has less hallucinations, was able to get more relevant results (e.g. "most recent article on root.cz" <- this pplx constantly fails) and felt a bit longer. though I am not sure when pplx made the change to longer responses, so that may no longer be valid. in length openai was the most impressive one, though I didnt have a change to try it personally. if we ignore everything else, I think it is:
pplx deep research < grok 3 deep search < openai deep research
but the important things is price and limits. pplx has a lot of uses (some maybe even free?), is quite fast and costs 20$/m, grok I think has not disclosed limits but cheapest with search on grok.com is 30$/m and oai is the whopping 200$/m at 100 searches per month and it takes a lot of time (like 10-20 minutes I think). so, as always, pretty much depends on user's use cases and money. I personally see grok.com as lacking too much features compared to openai and pplx, most likely cheating in benches (the famous cons@64) and missing API, so not possible to run regular benchmarks independently. only worthwhile feature seems to be the search (depth and X as one of data sources) with undisclosed limits, se either they will have (almost) unlimited high-quality voice2voice or can't charge more than oai and pplx. oh, and grok 3 is a less censored model, though oai seems to be on a more open path as well and pplx is fairly uncensored (at least I think R1 and sonar are)

[pliant widget]

set your default model to 4.5 in settings then use auto. disable search probably is the way to go its unstable as is

[left stream]

Not yet, what should I test it with

[pliant widget]

i aint reading all that but i already agree every deep serach integration is better than pplx's one

it seems super rushed

[left stream]

I definitely like 4.5. If you want to analyze someone’s soul use deep research. If you want to touch someone’s soul use 4.5 😂

[sweet dew]

Loooool

[pliant widget]

anyone can write a script to serach 10k articles of content which it seems to be doing. the real challenge is qualitative analysis of good arguments over autorities and erreneous stuff which it seems to absolutely not be doing

[sweet dew]

That is true

I tried using the deep research for a research I was already doing normally and yes it did give more content but the arguments seemed bad

Like it didn’t quite understand each article and the relationship as well as the thought

[pliant widget]

its night and day and the current deep reasearch models im comparing it too arent even good lol

[placid coyote]

pplx's deep research can be useful, but hallucinates a lot, probably because of the R1. is shortest from all 3 options, but is cheap and has many uses

[pliant widget]

very disappointed honestly. they are losing moat and instead of refining their product, moving to random avenues and investing time in boderline scummy tactics like the whole ccp thingy

just blatant lies all around, arguments with no bias, fear mongering yada yada

[placid coyote]

4.5 seems to be working fine for me. but it is VERY slow https://www.perplexity.ai/search/compare-deep-search-research-f-t45GYsRNSg6bk6Hl8N0VeQ

[pliant widget]

sadge

[pliant widget]

it stops after a certain amount of tokens i believe

[sweet dew]

Yeah pretty slow

[placid coyote]

nowhere near the hard limit. this was 1.6k tokens. hard max is around 4.5k (today sonnet 3.7 got cut off)

[pliant widget]

whats this about 3.7 sonnet?

it got rate limited?

[placid coyote]

response length is on pplx limited (resp. more than only by the api provider, that is 8k I think?)

[pliant widget]

yikes

was bound to happen honestly

[sweet dew]

Btw am sorry but how much is 1 k token approx

[placid coyote]

it was always

[smoky folio]

why does it have a japanese mini lesson 😭

[sweet dew]

Like in terms of characters

[placid coyote]

for en usually 0.75 of word = 1 token

[pliant widget]

about 750 words

[sweet dew]

ok thx

[pliant widget]

xD

what even

[shy moss]

Has perplexity launched 4.5 in its platform ?

[placid coyote]

so yeah, not a clean test. but 4.5 seems to give shorter responses and sonnet 3.7 wasn affected. similarly to sonnet 3.6, it was a pain to get anything beyond 1k tokens

[pliant widget]

it did but i wouldnt recommend using it for long conversations

[placid coyote]

yes, 4.5 is here, but possbily not on phones yet?

[pliant widget]

also 10 proopmts per day

[left stream]

Lol

[placid coyote]

to be frank, I was surprised they even added 4.5. it is veery pricey

[left stream]

Ikr, I can see why it’s so expensive now. It’s good. I’ve been using it for a while now

[pliant widget]

generate green texts with it almost guaranteed laughs

[sweet dew]

Don’t we get like 10 prompts?

[sweet dew]

For pro users

[pliant widget]

"> mfw i have no face" had me cackling man istg

[left stream]

Yes

[placid coyote]

jesus christ, that cant be right

Each query costs $3.00 (as calculated previously)
10 queries per day = $30.00 daily cost
30 days in a month = $900.00 monthly cost
it must be lying...
https://www.perplexity.ai/search/calculate-price-of-32k-in-toke-uTNAppRNR9e_OgFO_4iuNA

[left stream]

That was a good one 😭

[pliant widget]

hmmmm 4o seems to also be good at it

[left stream]

Japanese mini lesson is crazy

[pliant widget]

idk i need better anecdotes from people cause it clearly isn't winning any benchmarks

[distant aurora]

Need to compare it with open ai so run similar prompts on both.

[sweet dew]

thas crazyyyyy

[left stream]

4.5 is more emotionally intelligent, not analytically

[pliant widget]

im testing on 4chan green text outputs lol

[left stream]

It made me laugh ngl

I wonder how well 3.7 sonnet does

[left stream]

Bet give me a prompt and I’ll test it

[placid coyote]

4.5 writes great, but the price is just sad. you cant chat with it if you only get few prompts, so maybe for serious writing it is good?

[pliant widget]

my guess was openai made 4.5 to create synthetic data by turning it into a reasoning model for future models but knowledge cutoff is oct 23 so idk now

[pliant widget]

proably better, shes amazing

[pliant widget]

the 4o one was also funny ngl

[distant aurora]

I think it will always be limited to coding.

[pliant widget]

some would say even better

[placid coyote]

in some writing bench I think 4.5 wasnt the king, which surprised me, since it is monster model with so much training

[distant aurora]

Should I give something complex?

[pliant widget]

also karpathy sensei made a small survey and gpt 4 won over 4.5 on vibes lol

[placid coyote]

ah, here - sonnet beats 4.5 in EQ and R1 in writing https://x.com/monnef/status/1895844428360843455

mennof (@monnef) on X

Am I missing something about GPT-4.5's supposed "emotional intelligence"? I'm skeptical about mixing intelligence with emotions conceptually, but shouldn't a model with such massive training & computational resources at least dominate EQ & creative writing benchmarks? #gpt45 #ai

X · Mar 1, 2025, 2:31 PM

[pliant widget]

nah like give it a ridiculous scenario and ask to complete

[sweet dew]

sonnet 3.7?

do we get unlimited prompts for 3.7?

(pro users)

[pliant widget]

be me open source maintainer for 70 OSS
complete the above script like 4chan greentext

im asking claudiepie

[placid coyote]

though to be a bit fair that was a biased test. only certain kind of people follows karpathy (ai geeks, predominantly probably some type etc) and he selected questions which he though the 4.5 were worse. so thats at least two major biases

[placid coyote]

I think so, it is the normal pro model, so the 600 daily? but not the thinking one - I mean sonnet 3.7 non-thinking

[pliant widget]

he preferred the 4.5 outputs but the majority didn't. idk about the choices for his question i just saw the results

[sweet dew]

ahhh makes sense

[placid coyote]

oh, okay, then I guess might be more fair. but still introduces unknown level of bias. like sample of geeks following him on X is not average human

[pliant widget]

kinda bad. claudipie is too smart for silly jokes

[placid coyote]

wouldnt be surprised if even we here would be skewed towards introverted geeks, or even borderline aspergers or something like that. I know I am 😅

[left stream]

👍

[placid coyote]

sonnet used to be great at roleplay, but sometimes would need more preparation (almost felt like jailbreak)

[left stream]

I’ll test it rn forgot I have my claude 3.5 jailbreak

That was wild

[pliant widget]

he might be one of the smartest people ik cause i could one shot his 3hr long videos with barely any pre requisite knowledge. just one sitting pure bliss. talking about karpathy sensei. his opinion was not with the majority at all so...

[placid coyote]

sonnet writes great after few "roast it more" on pplx ^_^

[left stream]

4.5 is on the app rn

[pliant widget]

this was good. charge your phone silly goose

ok bye now

[ruby heart]

App crashes with this prompt

[placid coyote]

I got weird crash on R1 on web today. when I refreshed the reponse was empty, didnt even saw a model icon/label thingy

[regal berry]

Hello!
Why did you never add the DeepSeek V3 model which is much cheaper, but you did add the GPT 4.5 which is dozens of times more expensive? Thank you if someone from perplexity can answer me

[distant aurora]

Use r1 via blackbox ai.

[distant aurora]

V3 ~ 4o. Didn't serve any special purpose.

Added gpt 4.5 probably cuz of open ai hype.

Bro, o3 mini high is a completely different level beast.

Wish I had free access, lol.

[regal berry]

Hmm, that doesn't make sense, since there's the 4o and it's inferior to v3

[limpid oracle]

Yea

[ruby heart]

https://youtu.be/TM6q1F45boQ?si=4rBuaf-1Q-Z0KUT_

Just a warning to everyone

[left stream]

Wonder what type of reasoning monster can be made thanks to 4.5

O1 from 4o and o3 from o1

[dense sparrow]

Geez no kidding

[robust remnant]

does perplexity officially have a discord bot

[left stream]

I’m calling dibs that the next reasoning model built upon 4.5 is gonna be called o5

[distant aurora]

O 3.5 would sound weird.

Lol.

[sweet dew]

are we getting grok 3 grok 3 reason and grok deep search anytime soon on perplexity?

[left stream]

True lol

[distant aurora]

Okay, I tried a problem on the perplexity version of r1

It didn't match with the deepseek version of r1.

https://www.perplexity.ai/search/there-s-a-square-edfb-of-6cm-i-NBfkPiGpSoOjxpsbjSg1SA#0

[sweet dew]

lemme guess perplexity was a bit worse?

i have seen similar results with o3 mini as well

even 4o

not just r1

[distant aurora]

Actual r1 https://www.blackbox.ai/share/bfeea050-ed74-45bc-90c9-abb0cd900f90?mobile=true&model=DeepSeek-R1

[distant aurora]

Yes.

[sweet dew]

yeah oh and deep seek r1 now takes upto 50 documents

why is perplexity limited to 4?

[distant aurora]

If you are using documents, you might wanna try notebook lm.

[sweet dew]

oooh ok i'll try that

another good (chinese) model i tired is kimi ai

[distant aurora]

I haven't tried kimi yet. Only saw reels about it.

I have tried qwen.

[distant aurora]

O3 mini high solved this problem in one go without much thinking. https://sharetext.io/d931a669

The free o3 mini took way longer and thought more.

[sweet dew]

i didn't like qwen honestly

[distant aurora]

I didn't use it for anything complex yet.

[distant aurora]

Why not?

[sweet dew]

maybe i got the prompts wrong but the same prompt which i gave to r1 and o3mini, qwen performed much much worse

imo even kimi was better

[distant aurora]

Omw to try it

Okay, it solved my problem.

[distant aurora]

What was the prompt?

[sweet dew]

it was a coding based prompt

i was trying the hexagon with circle bouncing inside

with every different model

and honestly best was claude 3.5 (3.7 didn't exist then)

[distant aurora]

Oh.

I'm assuming you used the coder model then.

[teal furnace]

Does anyone using perplexity for coding? How does it compare with native models.

[distant aurora]

Native better if you ask me.

I tried making excel macros.

[teal furnace]

I'm thinking more about python.

[distant aurora]

But then again, depends on the length and complexity of the task I guess.

[distant aurora]

I think most people here would agree that native almost always performs better.

[teal furnace]

Thanks !

The issue was having to pay for multiple services instead of just 1. I guess I'll pick one of the "known good ones" and stick with it

[distant aurora]

Yeah. Fair.

You can access deepseek r1 for free.

So your options are limited to Claude or chatgpt I guess.

[sweet dew]

yes

[distant aurora]

I would suggest trying out/asking people for alternatives before any purchases.

[sweet dew]

imo if it's only code related try to go for claude

[teal furnace]

I'm already paying for Claude. I guess I'll still stick with it. Was wondering if switching to Perplexity in order to use different models but I guess better not.

[sweet dew]

it's the best in coding sector

[teal furnace]

Thanks bros.

[sweet dew]

why not give both of them a complex task

and compare results

[distant aurora]

Try qwen max. They even added thinking option.

[teal furnace]

The issue I'm having is that it seems perplexity doesn't remember the past very well. It fixes some issue but doesn't analyze the full project

[sweet dew]

yeah that is true

[sweet dew]

will try

[distant aurora]

You get 10 o3 mini uses for free per day on chatgpt.

[teal furnace]

So it fixes something while breaking something else

[sweet dew]

really???

[distant aurora]

Yes

[sweet dew]

i thought i got more than that

never actually run out

lol

[distant aurora]

Lmao. Fr?

[sweet dew]

yeah

maybe cuz i have 3 accounts for 3 different browsers

🤣

[distant aurora]

My friend told me something similar too but I ran out yesterday.

Or you can use deepseek r1 via blackbox ai.

It's fast and free.

Downside is that it doesn't save chat history.

So you can choose share link and save the link somewhere.

(groq has limited context window and open router was sluggish in my experience.)

[left stream]

I’m a fool for using deep research on this question when you already solved it 😭

[distant aurora]

Wait. I'll give you a more complex prompt later.

Something that perplexity failed in.

But it will be more about research and less about problem solving.

[left stream]

https://sharetext.io/be8ae4c4

[left stream]

Ok

[tame current]

Was that not one of the things they were saying they improved on? 😭 Thanks for the info btw 🙂

[left stream]

I can say definitely without a doubt, it beat 3.5 mini medium. For reasoning tasks, it’s best o1 or deep research (o3). If you’re trying to learn or study. 4.5 is good to use. It’s basically a very good explainer while the reasoning models are the best analyzers

[gilded finch]

What’s the best api for research

[wintry lintel]

Hi, I ve trouble in getting content from a website since I am unable to get all the info from the website because there are too many things on the website I want to obtain while you do not support to fetch retrieve contents from weblink. Is there a way I can getthat infos from website without having to spend a ton of times to do it nanually? Thanks
Đư

[dry kestrel]

Its not good

they only have the base version of o3 mini and claude 3.7

[patent cargo]

4.5 limit? 10 a month or day?

[dry kestrel]

day

[faint spoke]

perplexity frequently send me promotional emails about deep research, yet I still can't privilege my sources from the URLs quoted, I still get responses from Day 1 releases that are full of lies, and I still have to wait for the next day's release to get a response from them. I am about to cancel my subscription.

[naive sage]

Could you explain what these statements mean, I'm not sure I understand them.

1. "I still can't privilege my sources from the URLs quoted"?
2. "Day 1 releases that are full of lies, and I still have to wait for the next day's release to get a response from them"?

[amber lagoon]

im having trouble with perplexity enterprise.
i made a space and invited a contributor and made some threads to test. he made a theed also.
i can see his threads but he cant see my threads
is this a bug or is there a setting i can change to fix the problems

[jovial kayak]

I don't know if we're allowed to discuss Complexity, but I just tried out the Canvas plugin (Claude artifacts-inspired), and it's actually so cool. Anyone have some really complex (pun intended) prompt they used that worked?

[pallid elbow]

I don’t understand what you mean…

[amber lagoon]

i think the app on iphone i broken. spaces isnt syncing for all contributors

[fickle sorrel]

10x 4.5 ? perplexity getting poor?

u know ohw expensive that is

reading 40 sources...

sure input isnt 150 euro per million tokens but input is still 75 euro

[silk obsidian]

How long does it take to get a response from the pro support? Its been around 24 hours and im just getting AI responses.

[mossy pewter]

When Grok3?

[naive sage]

Likely very quickly after it gets an API, because that's required for integration. From https://x.ai/blog/grok-3:

Grok 3 API Coming Soon
In the coming weeks, we will release Grok 3 and Grok 3 mini via our API platform, offering access to both the standard and reasoning models

[brittle hill]

I'm experiencing an issue with the microphone not working in the Perplexity AI app on my Samsung Galaxy Tab 6A. The microphone works fine in other apps, but not in this one. I've tried restarting the device and reinstalling the app, but the problem persists. Can you please help me resolve this issue?

[tranquil orchid]

Hey @brittle hill!

If you find the original message helpful, please consider reacting to it with the :star: emoji. If the post is appreciated by the community and receives 5 stars, it will go to the https://discord.com/channels/1047197230748151888/1082806833938436228 channel and the post author will get the <@&1082034222778302614> role on Perplexity.

[placid coyote]

well, the price is kinda insane. that calculation assumes pplx has same settings as for sonnet (max in and out sizes) https://www.perplexity.ai/search/calculate-price-of-32k-in-toke-7GV3AF.zSOuXXL2PIdpUeg

[tiny viper]

I would have preferred if they had invested in Claude's thinking model honestly, given that GPT-4.5 doesn't bring anything new

Btw, Perplexity had any new updates in the last few hours?

[placid coyote]

arent they working on thinking sonnet? but I would guess it will have more strict limits, since while not as pricey as 4.5, it still isnt very cheap (isnt R1 like 2$ while sonnet 15$?)

[distant aurora]

He means that he wants a response based on the webpages he mentioned in the prompt.

But perplexity won't obey.

Does anyone here feel like gpt 4o is way more dumb than gpt 4 turbo?

https://chatgpt.com/share/67c416a0-1e1c-8013-8066-4a37eae3afb7

I remember using gpt 4 turbo to calculate import taxes.

We should have a help channel here. Or else people's queries will just get lost.

[vestal badge]

are there limits on deep research for pro users in a day ?

[naive sage]

I think you can use all of your pro searches on Deep Research queries if you choose, so it's the same limit (at least as I understand it - the limits and how they work could be a bit clearer, but the same criticism that clearer limits would be helpful applies to other providers like OpenAI/ChatGPT and Anthropic/Claude as well)

Questions like that are extremely common in here from people, which is always an indicator that it should probably be more clear.

[tiny viper]

That's why I was saying that instead of investing in a relatively weak LLM (in terms of quality-price), it would be better to invest in one of the more powerful models on the market. Though, I think PPLX integrated GPT-4.5 only because they probably have some collaboration with OpenAI to encourage its use, since it hasn’t received great public feedback due to recent benchmarks and its price.

[vestal badge]

i dont think there is limit on pro searches for pro users and if the same apply there is no limit on deep research ?

for pro subscribers

[tiny viper]

Nope, only sonar default it's unlimited, Deep Research (in theory) have 500 use and the other 600 use ( https://perplexity.ai/rest/user/settings )

[distant aurora]

Not bad if it actually works as intended.

[placid coyote]

yeah, probably have a bit cheaper oai api or via 3rd party partner. but the thing is, I believe there is a special anthropic api for the reasoning budget, so that may take some time to integrate and might be reason why they picked 4.5 first, since its easier to integrate and has bigger pr potential, since sonnet is primarily for programming and not sure if sonnet thinking is worth it (slightly more "intelligent" but costs several times more)

found it, on dubesor sonnet thinking cost 7 times more. not even sure, how big budget, but anthropic used 32k in benches so that could be the point of diminishing returns. if we would use current pplx numbers (max 4k tokens out) and add 32k thinking budget, then output could get 9 times more expensive. I don't think we are getting that with hundreds of daily uses 😅

[timid marten]

How to open this mode in library?

[glass yarrow]

Anyone from Bangalore??

[serene sand]

within a few business days

[silk obsidian]

Thats very smart of them, since I'm trying to refund a 200 dollar subscription and its about almost 3 days since they replied