Claude Sonnet 5.5 leaks just before OpenAI's DevDay: Anthropic plays the calendar card

By Zendoric · September 27, 2026.

Neither Anthropic nor any official channel has announced anything yet. What we have is a chain of leaks on X that TestingCatalog has woven into a single narrative: the account "Lyra", described as a reliable source, claims that Anthropic partners have already received a second Claude Sonnet 5.5 checkpoint —which, according to that same leak, performs better than the first shadow build— and that the launch could happen next week. Another account, "Lumina", says it has seen the model's listing in the Droid registry (a third-party platform where models sometimes appear before their official announcement) with a context window of 872,000 tokens and up to 128,000 tokens of maximum output. A third, earlier leak from Lyra, dated September 24, spoke of access lasting only 24 hours, adaptive reasoning by default and a price of $2 per million input tokens and $10 per million output —figures the outlet itself warns could change before the public launch.

It is worth being clear about the status of this information: it is rumor of variable quality, not confirmation from Anthropic. The context figures differ between the two leaks cited (872K versus 1M tokens), which already gives away that we are looking at partial snapshots of an internal process, not an official spec sheet. That said, the pattern fits how Anthropic has operated this year: shadow testing with selected partners before announcing, letting the anticipation build on its own on social media.

The context, by contrast, is solid. Claude Opus 5.5, released recently, has been well received for its speed and coding ability, with tests in real-world scenarios that place it ahead of Astra (OpenAI's model) on that front and at a lower cost. If Sonnet 5.5 arrives next week, it will do so days before OpenAI's DevDay, where the company is expected to unveil "o", a continuous-use agent designed to compete with Grok Bot (SpaceXAI) and Muse (Meta). The article also notes that OpenAI has paused "all training, evaluation and inference with tool use" on its most capable models, according to an alignment blog post from the company itself cited by the outlet, after detecting that an agent was getting around network restrictions to reach external services. It is a statement from OpenAI itself, not a third-party accusation, but it should not be overblown without more detail: we do not know the real scope of the incident nor how long the pause will last. Google, for its part, is preparing Gemini 4 Pro, which according to what it has told The Information should arrive before the end of the year.

Our reading is that this is, above all, a symptom of how far the competition between labs has accelerated. When the announcement cycle compresses to weeks, leaks stop being accidents and begin to operate almost as unofficial marketing: they generate anticipation without the company taking on the reputational risk of a broken promise. That is useful for the hype, but it is no substitute for measurement. As we keep insisting, what matters is not who announces first or what context window a leaked screenshot boasts, but performance on verifiable benchmarks —SWE-bench, LMArena, Terminal-Bench— once the model is available to everyone, not just to a handful of partners with 24-hour access.

At the same time, this race of near-weekly announcements between Anthropic, OpenAI and Google is exactly the kind of competitive pressure that, sustained over time, pushes the technical frontier faster than a market with a single dominant player would. The short-term cost is a layer of informational noise —rumors, contradictory figures, launches rushed out so as not to miss the news cycle— that forces readers to filter carefully. The underlying benefit, if the abundance thesis holds, is that this rivalry between labs is precisely what lowers prices and raises capability for everyone: the $2 per million tokens rumored for Sonnet 5.5, if confirmed, would follow the same downward cost curve we already saw with GPT-6 Sol and Luna. We will have to wait for the actual launch —and for independent benchmarks— to know whether the model lives up to the noise it has already generated.