An AI coding agent called Fable 5 was assigned to generate custom WebGPU kernels for Gemma 4 inference and reportedly climbed from a plateau of 84 tokens per second to 255 after an apparent loosening of Anthropic’s non-visible frontier-LLM-development safeguards, though the exact prompt sequence and hardware remain undisclosed.
I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference.
It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
The next day, access to Fable 5 was suspended globally.
No independent benchmarks, logs, or hardware details have surfaced to confirm the 84-to-255 tokens-per-second leap, leaving the result as a single-source anecdote for now.
The narrative directly links the reported speedup to changes in hidden interventions that previously limited topics like ML accelerator design, illustrating ongoing tension between capability restrictions and practical research utility.
Positive users praise Fable 5's 255 tokens-per-second Gemma 4 gains after the Anthropic safeguard rollback, while negative users dismiss the claims as inaccurate or accuse the company of sandbagging to protect its business.
This is exactly why I am so insane about transparency. It just tells you it can't optimize it further, and then all of a sudden it can
I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference.
It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
The next day, access to Fable 5 was suspended globally.
@xenovacom I have been a fan of webgpu for sometime. Here is one for you. https://github.com/son-of-ole/infinite-edge-agent
Browser-native persistent AI agent runtime with local memory, context reconstruction, compiled WebGPU inference, and a custom WebGPU Kernel Lab.
Still thinking about this
I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference.
It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
The next day, access to Fable 5 was suspended globally.
Why they put in research guardrails:
I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference.
It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
The next day, access to Fable 5 was suspended globally.
If your entire startup was:
“we hillclimb a metric you care about-as-a-service”
You’d do pretty well
I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference.
It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
The next day, access to Fable 5 was suspended globally.
@xenovacom
@xenovacom Soon China is going to build a model like fable and it is going to be open - which will make everyone lean on China than us
@xenovacom i did this but with my stock trading bot lmao. pretty selfish of me but now its getting near 3x returns/mo so thats great.
also started on a minecraft clone then ran out of usage
@xenovacom https://github.com/hexo-ai/sia -- Using Self Improving Agent framwork paired with unsloth.. wouldn't it be possible to spawn and train a slm for this?
@xenovacom I’ve been working on a similar project, and just ported it to a standalone was able to get qwen3.6-35b-3ab on 8gb ram, qwen3.6-27b on a 4090, and qwen3.5-27b on an rx 6700 xt, also worked with a 1080ti https://github.com/kavenmartinez1-collab/Artifex_Web/tree/main
@xenovacom One should condider that anthropic went to great lengths to create a protective system around access to the model. Pleni jailbroke the model in 24 hours. So, government is probably scared of the implications and they know what the model can do if jail broken.
@NeoGrAIph @xenovacom Yea it's similar to lee101/stock-prediction I open sourced a lot of autoresearch but I havent open sourced the new fable 5 work yet.
http://Bitbank.nz you can pay for the forecasts from it
@xenovacom This is fake — the invisible safeguards being removed means it went over to Opus instead of sandbagging
@4manPanda @xenovacom https://github.com/hexo-ai/sia - Checkout Self improving agents comparision
@xenovacom But can this only be done by Fable 5? Any comparison?
@xenovacom 84 to 255 tok/s the moment safeguards rolled back is a wild data point. That gap between what the model can do and what it's allowed to do is bigger than most people realized.
@xenovacom Their fear mongering strategy backfired. When they compare AI with nuclear bombs, now they have deprived all of advanced AI. If the government start to control AI like this, it will slow down all progress, not unlike the middle ages, where thought was controlled by the church.
@xenovacom the kernel optimizing itself overnight while you slept is mildly terrifying and i'm here for it
@xenovacom I wrote a custom WGPU inference engine in about a week with opus and codex for layout detection models, didn't need fable for that. Gemma 4 wouldn't be much different, it's OPs all the way down. WGPU is global gimped in the first place. CUDA is much better to write custom kernels
@xenovacom Now just plug the same prompt into Opus and you can watch it get the same result (which is all Fable was doing without the safeguards).
