Claude Sonnet 5.5: What the Opus 5.5 Leaker Says Now

Anthropic has said exactly one thing about Claude Sonnet 5.5, and it said it in someone else’s launch post. Announcing Claude Opus 5.5 on 22 September, the company wrote that “Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks”, with many of the same gains in performance, efficiency and safety. That makes this an unusual rumour cycle: the model’s existence is not in question, only its date, its specs and its price.

As of 28 September, none of those has appeared anywhere Anthropic controls. Its model table lists claude-opus-5-5 and claude-sonnet-5, with no Sonnet 5.5 identifier. There is no model page at anthropic.com/claude-sonnet-5-5, where Opus 5.5’s page lives, and the pricing table has no row for it. Sonnet 5 is still the current mid-tier model, at $2 per million input tokens and $10 per million output.

The leaker with a track record

What makes this round worth following is who is doing the talking. On 26 September the X account @lyraxana, which signs as “lyra”, posted that partners had received a second Sonnet 5.5 checkpoint “even better than the first one”, with release in the coming week. That is the same account that, on 20 September, put a codename, a launch date and a full price sheet on what turned out to be Opus 5.5. As we noted after that launch, it got the name, the day and every price right, down to the cache rates, while the leak chain around it had renamed the model twice.

One correct call is not a record, and lyra has no stated connection to Anthropic. But it is the first time this year that a single Anthropic leak source has a verified hit behind it, and that changes how its next claim should be weighed. According to TestingCatalog, lyra has since suggested Sonnet 5.5 could arrive “as soon as early next week, possibly before Monday morning ends”. Monday morning, 28 September, came and went without it.

What has leaked, claim by claim

Claim	Where it comes from	Status
Sonnet 5.5 and Haiku 5.5 ship “in the coming weeks”	Anthropic, 22 September	Confirmed
A second checkpoint is with partners; release this week	@lyraxana on X, 26 September	Unconfirmed
Partners got 24-hour access; launch 30 September or 1 October	X account Pankaj Kumar, via Startup Fortune	Unconfirmed
Context window of 872K to 1M tokens, 128K output	Factory’s Droid registry, via TestingCatalog and WinCentral	Unconfirmed
Context window of 2M tokens	Earlier X posts	Contradicts the registry figure
$2 input, $10 output, $0.20 cache reads per million tokens	TestingCatalog	Unconfirmed; same as Sonnet 5 today
Adaptive thinking on by default	TestingCatalog	Unconfirmed; matches Opus 5.5
Codename “Fennec”	X posts	Recycled from two earlier leaks

The most concrete of these is the Droid registry, the model list in Factory’s Droid coding agent. Testers cited by TestingCatalog say it carried a Sonnet 5.5 entry with an 872K-token context window and 128K of output, and WinCentral reports the identifier claude-sonnet-5-5 turning up in production files. A registry entry is the kind of artefact that settled Opus 5.5, but so far it exists only in screenshots and second-hand accounts. It also fits oddly with the rest of the line-up: Anthropic’s current Opus and Sonnet models, Sonnet 5 included, all offer 1M tokens, and 872K is not a figure any of them uses.

The 2M-token figure, which ran in earlier posts alongside the promise of “Fable 5.1-level performance” at Sonnet prices, is harder to square with anything. It would double Anthropic’s largest context window in a mid-tier model, and it comes from accounts with no hits behind them. When two leaks disagree by more than a factor of two on the same number, at least one of them is wrong, and the one attached to an artefact deserves more weight than the one attached to a wish list.

The price is the quiet detail

If the leaked sheet is right, Sonnet 5.5 would cost exactly what Sonnet 5 costs today. That would be a break from the Opus pattern: Opus 5.5 came in 20% below Opus 5 on both headline rates, at $4 and $20. A flat price is plausible for a mid-tier model that is already Anthropic’s cheapest route to agentic work, and a leaker who simply copied the current Sonnet 5 row would produce the same numbers. The price, in other words, is the one claim here that would look identical whether the leak is real or not.

The codename deserves the least weight of all. “Fennec” first surfaced in February, attached to a claude-sonnet-5@20260203 identifier in a Google Vertex AI log, and that checkpoint shipped two weeks later as Claude Sonnet 4.6. It came back in June for Sonnet 5, as we covered at the time. CellCog points out that it is now being attached to a third model. A codename that has been applied to three different releases tells you nothing about any of them.

How soon is plausible

Anthropic’s wording puts no date on it, but the history offers a range. Sonnet 5 arrived on 30 June, about four and a half months after Sonnet 4.6. A Sonnet 5.5 within two weeks of Opus 5.5 would be by far the fastest mid-tier turnaround of the Claude 5 generation, though “in the coming weeks”, written on 22 September, reads naturally as October. The competitive calendar points the same way. OpenAI holds its DevDay on 29 September, and GPT-6 Sol and Luna have spent a week undercutting everyone on price.

As with Opus 5.5, it will be settled by public artefacts rather than by posts: a claude-sonnet-5-5 entry on the Models API, a model page, a pricing row. Anthropic publishes all three on launch day and never before. Until then the fair summary is short. Sonnet 5.5 is real and coming, the one source with a hit behind it says this week, and the specs attached to it have not yet agreed with one another.