Fable 5.5 is coming. Here's how we know (and what would prove it).
Anthropic has not announced Claude Fable 5.5. Its model docs still list Fable 5.1, released September 1, as the newest Fable. And yet for the past week the AI corner of X has been treating Fable 5.5 like it's a few days away.
That's not just hype. There's a trail of receipts, and it looks a lot like the trail Opus 5.5 left in September, right before it launched. Here's every piece, how solid each one is, and what would actually settle it.
Receipt 1: it showed up on Microsoft Azure
On October 13, LiteLLM's automatic price-sync bot added a new entry to its public model list: azure_ai/claude-fable-5-5. LiteLLM is the open-source library a lot of developers use to call many AI models through one interface, and the bot pulls straight from cloud catalogs. This entry came from Azure's East US 2 region, with the status "GenerallyAvailable".
Alex Getman posted the diff on X that afternoon. It passed 326,000 views. TestingCatalog followed an hour later: "If this model drops today, it would be a new SOTA across many areas. Soon?"
How solid: the entry was real. You can read the commit yourself. But look at the numbers inside it: $10 per million input tokens, $50 per million output, 1M context, 128K output. That's exactly Fable 5.1's spec. It looks like a placeholder copied from the previous model, not a real price list. The entry was gone by October 16. So the name existed in Microsoft's catalog. The details don't tell us anything yet.
Receipt 2: partners reportedly have it
On October 15, the leaker Lyra posted: "Partners got access to the upcoming Claude Fable 5.5. No info on any release date this time but I assume next week." TestingCatalog repeated the report on October 17.
Why this one matters more than a random account: on September 20, Lyra wrote that Anthropic was "stealth testing Opus 5.5 (claude-opus-5-5) under the codename claude-wafer-eap which is planned to be released on Tuesday." Opus 5.5 launched that Tuesday, September 22. Same leaker, same kind of claim.
How solid: reported, from one source with a good record. "Next week" is Lyra's own guess, and the post says so.
Receipt 3: a codename, onigiri
Also on October 15, LuminaBench found a new codename in the frontend code of Artificial Analysis, a site that benchmarks AI models: claude-onigiri-eap. It came in four effort levels (low, medium, high, xhigh) and listed Opus 4.8 as its fallback, which is the pattern Fable and Opus tier models use. The entry disappeared within hours.
"EAP" is the suffix Anthropic's early-access codenames have used before (Opus 5.5 was claude-wafer-eap). Separately, people reading the Devin CLI found a string of other new codenames: Orzata, Boba, Vanilla, Odango, Verrine and Onigiri, with build dates from September 21 to October 10, per TestingCatalog.
How solid: reported. Lumina said "could be wrong" in the same post. Six codenames could also mean several models, not one.
Receipt 4: Opus users got a smarter model
This is where it started. Over the weekend of October 10-11, people using Opus 5.5 in Claude Code started posting outputs that didn't look like Opus 5.5.
Chetaslua posted an animation where "every frame" was drawn in JavaScript, from two prompts of "literally 4 lines". The post lists 2 hours, 27 million tokens and max thinking effort, and opens with "Claude Opus 5.5 is routing to a new model", adding that it "can be fable 5.5 or new opus". The post passed 509,000 views. The same day, Marmaduke shared a full 3D Blender scene built from one sentence over two hours, calling it "the single most impressive output I've got from a model so far."
TestingCatalog wrote it up on October 12, and a quick test went viral: ask "Do you know VoxelBench? Do not search." A model that knows it supposedly has fresher training data.
How solid: this is the weakest receipt. No screenshot carries server metadata saying which model answered. LuminaBench, one of the people tracking this most closely, said they didn't notice "too crazy of a difference." And the VoxelBench test proves less than it seems: VoxelBench has been public since September 2025, so plenty of current models can describe it.
The precedent: last time it took 4 days
Here's what makes the pattern interesting. Opus 5.5 went through the same steps:
- September 18: routed outputs that looked like a new model (TestingCatalog)
- September 20: Lyra names the model and the launch day
- September 22: Opus 5.5 launches
Fable 5.5 so far:
- October 11: routed outputs
- October 13: the Azure listing
- October 15: partner access and the onigiri codename
- Launch: not yet
Fable 5.1 took longer. Leaks pointed to "the end of August" and it shipped September 1, about two weeks after the first signs.
What the market says
Polymarket runs a market on when the next Fable model ships. When I read it on October 18, with about $585,000 traded:
- by October 21: 38.5%
- by October 22: 51.5%
- by October 24: 65.5%
- by October 31: 80%
Two caveats. The market counts any new Fable model (5.2 or later), not specifically 5.5. And it has been wrong already: bets on October 15, 16 and 17 all resolved No. A widely shared post predicted a Tuesday October 6 release, which also didn't happen. The October 24 odds were around 82% on October 16 before falling back to the mid-60s.
What would make it official
Leaks are fun, but four things settle it. Save this list for launch day:
- A launch post on anthropic.com/news. Every model launch gets one.
claude-fable-5-5in Claude's model list in the official docs, next toclaude-fable-5-1.- Real pricing. If it comes in at exactly $10 / $50, the Azure placeholder was right after all. If not, you'll know the leak was a copy.
- A system card with benchmark scores. No numbers have been published yet. Anything you see before this is a claim.
What to do this week
If you build on Claude, keep the model name in one config value, not scattered through your code. When a new model lands, you want to try it by changing one line, then compare it on your own tasks before switching anything important over.
And if someone shows you a "Fable 5.5 output" before launch, ask the boring question: how do you know which model answered?
Which of these receipts convinces you most? Tell me on X or LinkedIn.