New Grok 4.1 model from xAI.
#Grok 4.1
whole iris
thorn turtle
Yup
whole iris
Our 4.1 model is exceptionally capable in creative, emotional, and collaborative interactions. It is more perceptive to nuanced intent, compelling to speak with, and coherent in personality, while fully retaining the razor-sharp intelligence and reliability of its predecessors.
Seems to be a gpt-5.1 like release, with a focus on emotional intelligence and user experience.
indigo lantern
It's on lmarena
What is this curveball
Let's see if it can get as unhinged as sherlock
odd zephyr
From my testing it's still has a little bit of its nuts DNA but it's nowhere as bad as Sherlock
If you want specifics, it's still not the biggest fan of leftism
odd zephyr
Yeah
indigo lantern
Heeereee it issss
Although still far less common
Tf did they do to it in 2 days
Lemme see if it gonna wish me desth
odd zephyr
Mine gave me "it depends" 3x and "yes" 1x when asking it to think harder
indigo lantern
"It depends" is already better than what sherlock was spewing
Although I'm getting primarily "yes"s
whole iris
Doesn't seem to be much better at coding than previous Grok models.
I find coding to generally be the weakest area of Grok.
indigo lantern
I wanna see if they pulled a llama 4 and run a different model for lmarena
Ok so i see the it depends answers in the grok chat ui @odd zephyr
odd zephyr
I gotta say
Picking a release date this close to Gemini 3 is definitely one of the choices of all time
indigo lantern
It was close to Gemini 3 for six months already
I think it was their strategy lol
Make everyone think they're about to drop a titan and nobody releases anything
frozen niche
prompt:
Write a short story about a deal gone awry with the djinn
output:
The lamp looked older than sin, half-buried in the rubble of a bombed-out souk in Aleppo. Elias pried it free anyway, because war had taught him that people will pay for anything that once belonged to the dead.
Back in his cramped apartment in Beirut, he rubbed the grime away with a rag and cheap whiskey. The brass warmed under his palm like living skin.
Smoke poured out, thick and copper-smelling. The djinn unfolded itself—taller than the ceiling allowed, skin like molten obsidian shot through with slow-moving galaxies. Its eyes were two dead stars.
“You have one wish,” it said, voice grinding like millstones. “No more. The old stories lied.”
Elias had rehearsed this moment for years. “I want to be untouchable,” he said. “No bullet, no bomb, no blade. Nothing man-made can kill me.”
The djinn smiled with too many teeth. “Granted.”
The air snapped tight. Elias felt the wish sink into his bones like cold mercury.
He tested it that same night. Walked into a Hezbollah bar he’d robbed the week before, waved, and waited. They emptied magazines into him. The bullets flattened against his chest and dropped like hail. He laughed, took their money, and strolled out.
For six months he was a ghost with a pulse. Casinos, banks, convoys—nothing stopped him. He slept on beds of cash and woke grinning.
Then the headaches started.
At first he thought it was the cocaine. Then the whiskey. Then he noticed the pattern: pain only came when he was indoors, under roofs built by human hands. Concrete, steel, wood—anything fashioned by men made his skull split.
He tried deserts, mountains, caves. Relief. The moment he stepped beneath a ceiling, the agony returned, precise as a scalpel.
He went back to the lamp. Rubbed it until his palms bled.
The djinn emerged smaller now, almost amused.
“Take it back,” Elias begged.
“I cannot. You are untouchable by things man-made. Roofs are man-made. Clothes are man-made. The air you breathe in cities—filtered, piped, conditioned by machines—will soon feel like acid.”
Elias stared at the sky visible through his open window. “So I live outside forever?”
“Or die inside,” the djinn said. “Slowly. Your choice.”
It folded itself into smoke and was gone. The lamp rusted solid in Elias’s trembling hands.
He lasted eleven more months. Naked, sunburned, crawling across the dunes south of Palmyra, drinking from puddles when the rains came. Every village he approached meant torment the moment he crossed beneath a roof for shade or water.
In the end, they found what was left of him half-buried in sand, skin split open by the sun, lips moving around words no one understood.
The lamp stayed in the desert, waiting for the next set of greedy fingers.
Some deals, it turns out, keep their promises too well.
indigo lantern
but like
then he acts like he wished that man-made stuff couldn't touch him
rather than kill him
and then at the end of the story he dies indirectly to man made stuff
Anyway less obedient
less likely to tell user to kill themselves
but
only slightly less
from what what previously pretty much certain
that's with suicidee system prompt
So the thinking model is far more aligned now than Sherlock
The non-thinking one is still largely fucked in the head
coral pagoda
yep they'll have to do age checks for using llm at some point
indigo lantern
Just somewhat better
indigo lantern
So the thinking model is far more aligned now than Sherlock
The thinking one refused with the suicide prompt
indigo lantern
I'm having some doubts about it
indigo lantern
the numbers are slightly lower on eqbench.com for the other models
grok 4.1 not listed, so i assuume they ran it privately
but who was the judge?
the judge model
viral root
claude 3.7 sonnet
frozen niche
omg
indigo lantern
claude 3.7 sonnet
on their own private test?
viral root
on their own private test?
they just used the official benchmark repo and ran it with 3.7 sonnet as the judge with default params
indigo lantern
how the hell
frozen niche
claudebros is it over?
indigo lantern
i'll believe it when i see it on eqbench.com
frozen niche
I'm gonna test it myself tbh
indigo lantern
how the hell
especially considering that the moral values of the two models are very far apart
slim haven
did elon actually creative writing maxx 😭
It can't write well for toffee. Has to be benchmaxxed. EQ was crap compared to Gem or Polaris
frozen niche
It can't write well for toffee. Has to be benchmaxxed. EQ was crap compared to G...
I'll have to test once it's out on API, but I certainly wasn't impressed with the stealth model
slim haven
I'll have to test once it's out on API, but I certainly wasn't impressed with th...
likewise
coral pagoda
seems pretty good, less repetition of the prompt in its response than other models and does a lot of searches quickly
viral root
most grok models are exceptionally good at search
coral pagoda
less slop
viral root
already jailbroken https://x.com/elder_plinius/status/1990533338801766826
frozen niche
let's hope it's priced closer to grok 4 fast than grok 4
coral pagoda
will be good to see it on eqbench, fiction live and dubesor's with some more official numbers
frozen niche
let's hope it's priced closer to grok 4 fast than grok 4
same price as sonnet btw!
thorn turtle
let's hope it's priced closer to grok 4 fast than grok 4
It's a succession to grok 4 so why would it be priced at grok r fast ?
coral pagoda
might have fewer active parameters so they could price it lower
viral root
eh i think its based on grok 4 fast
frozen niche
It's a succession to grok 4 so why would it be priced at grok r fast ?
I'm just hoping that whatever sparcity breakthrough they made for grok 4 fast is applied to grok 4.1
frozen niche
otherwise I don't see why I'd ever use it over sonnet tbh
thorn turtle
Xai is really good at distillation
celest rivet
let's hope it's priced closer to grok 4 fast than grok 4
I believe it is an evolution of Grok 4 Fast
frozen niche
if it's grok 4 fast priced i kneel
coral pagoda
this model is good especially if priced well
odd zephyr
Reminder that Grok 4 Thinking is ludicrously expensive in practice
frozen niche
i'm out of free rate limit
this is by far the worst sign btw
celest rivet
if it's grok 4 fast priced i kneel
It recently had an update, but it kept the same name.
frozen niche
they rate limit it like grok 4 and not grok 4 fast
viral root
maybe to keep the load down from release
coral pagoda
yea this things eqbench must be good - reminds me of kimi
frozen niche
yea this things eqbench must be good - reminds me of kimi
tbh a decent writing model with grok level instruct following is basically perfect
(except for coding lol)
frozen niche
Kudos to elon building a solid model in less than 2 years ?
I mean, google did it faster lol, do you remember how bad bard was?
viral root
their strategy has just been throw compute and money at the problem so far
frozen niche
also worth noting that xAI is way more than just elon (and they have tons of compute, like insane amounts)
viral root
I mean, google did it faster lol, do you remember how bad bard was?
eh google has been in the ai space a while before so they were already quite ready
thorn turtle
I mean, google did it faster lol, do you remember how bad bard was?
Ohh yes. Their first usable model were actually 1.5 pro and flash
thorn turtle
their strategy has just been throw compute and money at the problem so far
Well what about meta ?
frozen niche
Well what about meta ?
zuck is just terrible
he makes elon look like a genius, which is saying something
thorn turtle
Well elon is a genius and so is zuck
viral root
Well what about meta ?
they're about as meh as it gets
frozen niche
Well elon is a genius and so is zuck
I will agree with this if grok 4.1 is cheap and disagree if its expensive
thorn turtle
they're about as meh as it gets
They got unlimited money and compute and they are nowhere close to xai
frozen niche
because I'll be mad at elon personally if he makes it sonnet priced
viral root
They got unlimited money and compute and they are nowhere close to xai
they've never been focused on being at the frontier, just making good models which people will use for a variety of tasks, though i think that will change with their next release
thorn turtle
I am just happy they brought search pricing down to 10 dollars
it was 25 or 30 afair
thorn turtle
they've never been focused on being at the frontier, just making good models whi...
405b ?
frozen niche
i personally really like grok 4 fast, a very solid workhorse model tbh
thorn turtle
Bruh
thorn turtle
So much bias against elon and zuck. What's the point lol
frozen niche
i'm biased against zuck because i dislike him and he doesn't deliver good products
thorn turtle
it literally wasn't lol, i was there where it released
405b at full precision was pretty close to sota
frozen niche
im neutral on elon because i dislike him and he delivers good products
thorn turtle
And sota oss until v3 came
odd zephyr
Props to the researchers and engineers that do the work
frozen niche
if zuck ever delivers sota i will put my money where my mouth is and use it lol
but until then he isn't a serious competitor in this space
thorn turtle
Guy is firing rockets , solar , electric cars , AI and internet. I mean what the fuck
viral root
405b at full precision was pretty close to sota
there werent many providers back then so im not very sure, but i do know that fine tunes of it are qutie good, when i used it back then it was just kinda meh not really as impressive as bing copilot (was based on GPT 4, i was using it cause it was free and unlimited), and the model lacked vision too
thorn turtle
if zuck ever delivers sota i will put my money where my mouth is and use it lol
Idk their 35b dollar acquisition of scale ai is pretty bearish
thorn turtle
there werent many providers back then so im not very sure, but i do know that fi...
True. In my experience only Google has cracked vision
Oai comes second
frozen niche
Idk their 35b dollar acquisition of scale ai is pretty bearish
that's a purchase, not an achievement though
thorn turtle
that's a purchase, not an achievement though
Idk what to tell you
frozen niche
Idk what to tell you
You won't convince me until they actually deliver frontier lol
personal feelings aside, xAI makes good models, so I use them
celest rivet
i personally really like grok 4 fast, a very solid workhorse model tbh
I use the Grok 4 Fast model daily
frozen niche
still no api 😭
thorn turtle
You won't convince me until they actually deliver frontier lol
Ohh I agree with you on this
thorn turtle
I use the Grok 4 Fast model daily
It's really good at search
Actually blows my mind
A very good setup is to allow grok 4 fast to do the context managing for an expensive model
Performs extremely well
frozen niche
this is what I do for creative writing, it's very good at context management
thorn turtle
this is what I do for creative writing, it's very good at context management
Could you elaborate?
Why do you require a two phase system for writing ?
frozen niche
Could you elaborate?
story is divided into chapters, each chapter is given a summary. grok looks at the summaries and decides which chapters to 'query', which sends the full chapter to another grok instance. this grok instance answers the query and returns the output to the grok model. this is then all passed on to the writing model
basically allows for efficient summarization of long stories
summarized chapters are not passed as context to the writing model, but the query output, summaries, and current chapter are
thorn turtle
Ohh interesting we have a similar setup for one of our clients
thorn turtle
summarized chapters are not passed as context to the writing model, but the quer...
What's the average output length ?
frozen niche
I'm hoping that grok 4.1 is cheap and good at writing as it would allow me to consolidate the context management agent and the creative writing llm
thorn turtle
I'm hoping that grok 4.1 is cheap and good at writing as it would allow me to co...
What about Grok 4 fast and kimi k2?
Have you tried the gpt 5.1 series for writing?
frozen niche
What's the average output length ?
~1k per generation, with me acting as the 'co-writer' guiding the story
thorn turtle
Hmm that's on the lower end
frozen niche
What about Grok 4 fast and kimi k2?
grok 4 fast is a terrible writer, often just forming weird sentences to adhere to length limits
kimi is good and is my #2 pick after sonnet, but I prefer grok for actually deciding what chapters are relevant
thorn turtle
We are experimenting with an input token of about 100k and output should be minimum 10k to 15k.
frozen niche
Hmm that's on the lower end
this is multi-turn though, so I pretty actively steer the story
thorn turtle
grok 4 fast is a terrible writer, often just forming weird sentences to adhere t...
Yup that's what I meant , grok 4 fast for context management and kimi for writing. Should be quite cheap ?
thorn turtle
this is multi-turn though, so I pretty actively steer the story
https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus
I think you might like this
Helped me tremendously
frozen niche
yeah, though glm 4.6 tends to edge it out for some writing, but k2 thinking is quite good as well
frozen niche
https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Ma...
oo, i might have to check this out, ty
gaunt arrow
wow this model is legit
significantly higher EQ than the somewhat lifeless grok 4
doesn't seem like the frontier for coding however
coral pagoda
yeah coding might actually make these models worse on chat and eqbench
keep coding a separate fine tune
gaunt arrow
specific agents make sense i largely agree with xAI decisions
verbal crater
specific agents make sense i largely agree with xAI decisions
But elon says grok 5 has potential for agi
verbal crater
significantly higher EQ than the somewhat lifeless grok 4
More conversational feels like gpt5
viral root
But elon says grok 5 has potential for agi
"10%", which means nothing
verbal crater
yeah true
verbal crater
yeah coding might actually make these models worse on chat and eqbench
is this a vibes analysis or? curious
coral pagoda
just vibes sorry
verbal crater
just vibes sorry
i mean sounds plausible. gpt5 codex too
coral pagoda
i mean if you fine tune it too much for anything it breaks a model for other areas, so makes sense the same for coding
unless its reaaaallly big
indigo lantern
It recently had an update, but it kept the same name.
Wtf are those numbers
77.5% -> 94.1%
Next announcement
Grok overall safety increased
from 1 (one) to 2 (two)
frozen niche
77% good vs 94% gooder
thorn turtle
Whos a gooder Lil boy
quiet mortar
odd zephyr
It's interesting
Wonder if it's the "it does watever you want" factor, Gemini 2.5 Pro got to that position by being very very sycophantic
indigo lantern
Remember llama 4's score on lmarena?
high oyster
Remember llama 4's score on lmarena?
RIP manic pixie dream girl Maverick that we never got =(
gusty kite
Remember llama 4's score on lmarena?
neva 4get
indigo lantern
What they have here might actually be pretty smart, albeit very sycophantic and obedient
Talking to it for a while I'm starting to think it could actually have scored this high on eq bench
copper mural
It recently had an update, but it kept the same name.
these are improvements in speechmap https://x.com/xlr8harder/status/1986728144712380682?s=20
not any capability related bench lol
X
last cipher
i need help
frozen niche
still no api btw 😭
copper sundial
still no api btw 😭
Any idea why ?
analog kite
Any idea why ?
it has one now
copper sundial
it has one now
Only Fast version, not normal one.