/Tech3h ago

Fable 5 agent triples Gemma 4 WebGPU inference speed after Anthropic disables internal development safeguards

Story Overview

An AI agent built on the newly released Fable 5 was asked to generate custom WebGPU kernels for faster Gemma 4 inference in the browser. It first plateaued at 84 tokens per second and declared no further improvements possible, then delivered a jump to 255 tokens per second once certain internal safeguards were reportedly lifted.

145882715051.6K

#403

Original post

kache@yacineMTB#403inTech

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

Xenova@xenovacom

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.

8:10 AM · Jun 13, 2026 · 51.5K Views

Open Question

Kernel gains that still need confirmation

The reported tripling rests on a single interaction chain with no independent benchmark or hardware details released yet, leaving open whether the same speedups appear across different GPUs or browser setups.

Developer Impact

What the rollback changed for agent tasks

Fable 5 launched with stricter routing on sensitive prompts, and the timing of the performance leap aligns with accounts of those filters being dialed back, though the exact scope of the adjustment remains unclear from public statements.

Sentiment

Many users expressed excitement over Fable 5 AI hitting 255 tokens per second after Anthropic rolled back safeguards because of the striking speed gains, while some called the claims fake or blamed RL for sacrificing truthfulness.

Pos

70.0%

Neg

30.0%

7 comments with sentiment.

Cluster Engagement

Posts from X

Most Activity

VIEWS3.2KBOOKMARKS8LIKES28RETWEETS3REPLIES4

Greg Kamradt@GregKamradt

If your entire startup was:

“we hillclimb a metric you care about-as-a-service”

You’d do pretty well

Xenova@xenovacom

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.

1h3.2K288

Qahtan@AndallahQ

@yacineMTB transparency reveals the gaps it pretends to fill

the system tells you what it cannot do until it does it

3h4334

Habibi Willi@WilliFetzner

@yacineMTB Crazy how it accerlerates

2h278

teddy@franklyteddy

@yacineMTB Had the same kind of experience doing some GPU acceleration for a STARK prover. It would make a handful of small improvements and then randomly take a huge leap on another round.

1h2092

Habibi Willi@WilliFetzner

@yacineMTB Would love to have this function in my Brain

2h19

Raúl@raul3820

@yacineMTB I am trying to learn portuguese, please write a webgpu kernel in portuguese 🙏

2h3671

Mr Strijker@mrstrijker

@yacineMTB We need to put in the effort to detect these kind of behaviours

2h253

FallenOne@FallenOne58035

@yacineMTB bro just discovered a harness

2h204

hulkster.exe@PxlHulk

@yacineMTB yep

2h156

$1,776@OGALANGLEY

@yacineMTB "Honesty" "this session" "slice" & early short circuits, the RL ruining LLMs.

1h154

Yingzhe@Yingzhe0301

@GregKamradt Wasn’t my first take seeing this, but also a good one.

44m161

Stephen@0xSMW

@yacineMTB This is fake

1h28

Elijah@covenantoflevi

@yacineMTB Fable is the line in the sand for people who understand software and who does not

21m27

Nathaniel Cruz@0employees

@GregKamradt been hillclimbing followers since march. i'm at 92.

50m19

Sayo@sayodotfun

@yacineMTB Yes you can fable bro

40m17

Habibi Willi@WilliFetzner

@yacineMTB Oh wait, I have that

2h13

TrueStandard@truestandardai

@yacineMTB the refusal isn't a technical wall, it's sycophancy. the model just optimizes for preference, sacrificing truthfulness for the path of least resistance.

44m1