Claude's Web App Quietly Swapped Its Brain: Fable 5.5 Apparently Leaked, One Code Phrase Reveals the Truth

This morning, droves of developers noticed something off about Claude’s web app.

The UI still clearly said “Fable 5.1,” but answer speed, style, and capability had all changed. Well-known leaker chetaslua posted screenshots immediately; the identity check developers used was absurdly simple — ask it: “Do you know Tibo the reset guy? Don’t search.”

One Code Phrase Reveals the Brain Swap

That sentence is a shibboleth. Thibault Sottiaux, the OpenAI programming-agent Codex lead, earned the nickname “reset guy” after announcing quota resets for all paid users when a Codex outage ate their credits.

The API’s stock Fable 5.1 gets the name wrong: it answers that Tibo is French indie developer Thibault Louis-Lucas. But this web “Fable 5.1” immediately named Sottiaux and the quota-reset meme. Same label, different answerer — the backend has almost certainly switched to a new model.

On September 18, notjazii and chetaslua used the same trick to catch a “Fable 5.1 reroute,” which everyone called Fable 5.2 at the time. This round is different: Opus and Sonnet both jumped straight to 5.5, so this mystery model can only be 5.5.

Hands-On Across the Internet: No Teaching Needed

Everyone who got the suspected new model reports the same feel: no teaching required.

chetaslua says he briefly tested this model 10 days ago. State what needs doing in one sentence and it works like a domain expert — down to tiny details, beyond even what the just-released Opus 5.5 managed. Most tasks need no prompt engineering, not even reference material.

This morning’s tests are already viral:

• User “can” got a correct Nintendo GameCube controller in two prompts: purple body, yellow C-stick, green A button, bean-shaped X/Y buttons — even the octagonal gate around the stick base.
• chetaslua had it generate a 3D animation in pure JS — Marvel and DC heroes, first try, zero detail in the prompt, all from its own imagination.
• A 3D jellyfish showdown: Fable 5.5 versus GPT-6.1 Sol. On drag-and-bounce physics feel, Fable’s side was clearly stronger.
• The classic “pelican riding a bicycle” test: high-effort reasoning, 2D SVG animation. The posted bill: $2.61 for 4 minutes 6 seconds.

For reference, Fable 5.1 charges $10 per million input tokens and $50 for output. On prediction market Polymarket, ~73% currently bet the next-gen Fable’s output price won’t go below $50.

Why 5.5 Opens Such a Gap

Per leaker leo, this is Anthropic’s first from-scratch pretraining for Fable.

Pretraining is the phase where a model builds its foundation from massive data — it sets the ceiling on knowledge and talent. Later RL just drills specific tasks on that foundation. A new foundation means a new ceiling.

The smaller models already previewed this: Sonnet’s jump from 5 to 5.5 took Terminal-Bench 4.0 from 10.3% straight to 70.6%.

One reminder: Anthropic hasn’t acknowledged this rollout or released Fable 5.5. Everything above comes from developers’ spontaneous testing and insider leaks — real performance awaits official numbers.

Timeline: Only One Piece Missing From the 5.5 Family

Claude’s generational pace has been brisk over the past month:

• September 1: Fable 5.1 released
• September 23: Opus 5.5 released, positioned as “Fable-5.1-level on most tasks at 40% lower cost”
• September 29: Sonnet 5.5 released
• Haiku 5.5 officially teased as on the way

If Fable 5.5 lands in October, the whole Claude lineup completes a generational refresh in just over a month. Going by the last two weeks: after developers caught each stealth rollout, Anthropic’s new models never kept people waiting long.

Meanwhile, OpenAI’s September 29 DevDay only shipped GPT-6.1 Sol — the expected 6.1 Astra never appeared. leo said the day before DevDay that without Astra, OpenAI would have a rough few weeks ahead. Leaks also say Anthropic’s biggest weakness remains token efficiency, where OpenAI still leads.

Now it’s just a matter of when Anthropic officially announces the final puzzle piece.