#Synthetic.new

19 messages · Page 1 of 1 (latest)

shell lava
#

https://synthetic.new/landing/home

$20 / mo for 135 messages every five hours
$60 / mo for 1,350 messages every five hours

Many of the popular models (a sample):

  • deepseek-ai/DeepSeek-V3.1
  • deepseek-ai/DeepSeek-V3.1-Terminus
  • moonshotai/Kimi-K2-Instruct-0905
  • openai/gpt-oss-120b
  • Qwen/Qwen3-235B-A22B-Instruct-2507
  • Qwen/Qwen3-235B-A22B-Thinking-2507
  • Qwen/Qwen3-Coder-480B-A35B-Instruct
  • zai-org/GLM-4.5
  • zai-org/GLM-4.6
  • embedding included for free
  • Supports custom models too on per-minute rates.
Screenshot_2025-10-08_at_7.10.22_PM.png
Synthetic | Run LLMs, privately

Chat with open-source models privately

quiet hornet
#

Mandatory 0.01$/Mtoken for the pay as u go embedding model price
For people who only want a super cheap embedding API

shell lava
#

@wheat thunder

I’m pretty pleased with Synthetic. Super responsive team. Wide variety of models for $20/mo. Their GLM 4.6 is quick - around 50-90 TPS. I also use terminus and qwen3-coder

wheat thunder
#

How fast is there Terminus and Kimi K2?

#

I currently use NanoGPT as my other provider, and they are okay but kind of slow to be honest. Broader model support than Synthetic so that's something

fringe goblet
#

this looks nice 🤔

shell lava
#

Terminus is solid for speed - not 1000tk/s or anything but I’m never waiting for it. I don’t use kimi at all.

Nano def has broader model support, but also backends to chutes for a lot of stuff so it’s spotty and the tool calling isn’t great.

Synthetic really spends a ton of time getting a few models right.

#

I’m a bit unique in that I still have more kilo credits than I know what to do with, but once those run out I think the majority of my spend will be with Synthetic on GLM and Terminus, and occasional spend with requesty or openrouter for the enlarger SOTA models like codex or Claude

#

They also have a super active discord with the founders

raw pawn
shell lava
raw pawn
# shell lava Code generation.

Oh interesting, pretty good code gen model? I've yet to use it. Need to do some testing on things other than Qwen, maybe I'll check out Terminus.

shell lava
#

It’s pretty solid. I want to do some work with 3.2 next

shell lava
#

Good news! The Synthetic team just pushed a fix to help them recognize Kilo tool calls the same as they do for Claude Code.

So now if you have JSON tool calling enabled with synthetic, your typed request in a sequence counts for 1, but all the tool calls initiated by Kilo count for 0.05 (like they do for Claude Code).

If you have good cohesive prompts that allow the model to work independently, this change will likely make your subscription go 5-10x further for the same workload

golden harness
shell lava
golden harness
# shell lava I implemented this today: https://github.com/mcowger/priceServer/blob/main/price...

sick. so, forgive the ignorance/lack of advanced research... my experience with chutes and hamfisted solutions the past few months has me a little jaded. where does synethetic sit in the grand scheme of things as far as the individual models and their ability to perform as expected?

do models such as kimi k2 thinking, glm 4.6, etc think/reason as expected with the provider as is? i understand any of this could be both kilo or provider limitations, so just wanted to mostly know what to expect.

shell lava
#

They don’t offer all that many self hosted models - not nearly like chutes does. I think currently it’s M2, GLM, K2-Thinking?

But the self hosted ones are done really really well. The 2 founders are super responsive to implementation bugs, usually fixing stuff within an hour or two. Super responsive on discord.

In at least 1 case, they are better than even the model lab themselves (GLM / Z.ai).

When I develop the JSON tooling stuff for kilo, there’s only 2 providers I trust to believe they have the right implementation, and that’s Synthetic and DeepInfra.