A model identifier called claude-sonnet-5-5 has turned up inside a third-party coding app, and developers say a few Claude Code users are already being routed to an unannounced model. Anthropic hasn't confirmed any of it, so treat what follows as a leak, not a launch.
When a new Claude model is close to shipping, the hints usually show up before the announcement. This time the clue is a single line in a client app's internal settings. On September 27, 2026, a developer posting as Lumina reported that the string claude-sonnet-5-5 appears in Factory Droid version 0.228.0, a coding tool that connects to Claude, tucked behind a feature flag while the app's public model list still stops at Claude Sonnet 5.
That's a meaningful place to find a name. Client registries are where vendors park real endpoint identifiers before the marketing switch gets flipped. It still isn't a release. Anthropic hasn't posted pricing, benchmarks, or a model card for Sonnet 5.5, and the whole thing rests on a screenshot and a developer's thread.
What the Claude Sonnet 5.5 leak actually shows
The claim has three parts, and it helps to keep them separate. One, the artifact comes from Factory Droid 0.228.0, a specific build of a third-party client, not from Anthropic's own servers. Two, what was found is an internal model registry, a file that maps readable names to provider slugs. Three, the slug in question is claude-sonnet-5-5, sitting next to other identifiers including claude-opus-5-5, which is already a live public model.
The circulating screenshot shows a fragment of that registry, with assignment lines pairing keys to quoted model strings. The Sonnet 5.5 entry is the one highlighted. The same frame lists other slugs that could be placeholders, future products, or unrelated tests, which is why registry dumps are noisy. One credible Anthropic-shaped name is worth noticing. It's not proof that everything in the file will ship.
The second half of the leak is about distribution rather than naming. The slug sits behind a feature flag, so the client knows how to ask for it while most installs shouldn't show it. Meanwhile Factory's public picker still lists Sonnet 5. A hidden name plus an unchanged public list is what people mean when they say a model looks staged.
Is Sonnet 5.5 in Claude Code greyscale testing?
Lumina added that Sonnet 5.5 is in greyscale testing, which usually means the company routes a small share of requests to the new model, or turns it on for allow-listed accounts, while everyone else stays on the current version. Replies on the thread follow a pattern seen in earlier Claude rollouts: a handful of Claude Code users claim they're already hitting the new model if they're lucky.
Those reports are anecdotal. There's no way from the outside to confirm who's in that group, or whether the model name in someone's interface matches production weights rather than a test endpoint. If you're not already in a test bucket, digging for hidden flags isn't a supported way to get early access, and it's a poor basis for any team decision.
Hearsay, for now.
That matters more than the slug itself.
Anthropic did say something official earlier. When it launched Claude Opus 5.5 on September 22, it stated that Sonnet 5.5 and Haiku 5.5 would follow in the coming weeks, carrying the same performance, efficiency, and safety improvements. That promise is the floor. Everything since is a question of how soon and what the identifier will be.
Why Sonnet 5.5 matters more than a flagship upgrade
Sonnet is the tier most developers actually spend money on. Claude Sonnet 5 has been the default workhorse since June 2026, and it sits at the middle of the lineup where the bulk of everyday coding and assistant traffic runs. A flagship model like Opus 5.5 raises the ceiling, but a Sonnet refresh changes the tool most people use daily.
If Sonnet 5.5 follows the shape of Opus 5.5, the interesting number won't be a leaderboard screenshot. It'll be the cost per completed task in Claude Code at a normal cache hit rate. Opus 5.5's launch leaned heavily on efficiency and price per task as much as raw capability, and developers are watching for the same pattern in the mid-tier.
That framing lines up with how the past two model refreshes have gone. Reliability on long agent runs and the bill at the end of the month often matter more to teams than a few points on a benchmark. A mid-tier model that holds flagship-quality coding closer to mid-tier pricing would reset the comparison that developers actually shop on.
Not everyone agrees Sonnet 5.5 is the point
The thread drew a familiar pushback. One reply argued that model upgrades barely matter anymore, because Sonnet 5 already writes better code than most reviewers check, so the real bottleneck is prompts, not weights. That's a fair point for teams with loose task definitions and light review.
It's less true for agents that run long loops, where small gains in reliability and cost add up. That's the group a Sonnet 5.5 refresh would target, and it's why the efficiency angle carries more weight than a headline benchmark.
Another reply guessed that Anthropic might ship on the Monday or Tuesday around OpenAI's DevDay, to grab some of the news cycle. That's a narrative, not evidence. Vendors do time announcements for competitive air cover, but registry leaks have also surfaced weeks before real launches. Don't plan around a guess.
What to do while Sonnet 5.5 stays unconfirmed
The safe move is simple: stay on documented model IDs until Anthropic publishes pricing and documentation. Chasing a feature flag in a client app isn't a supported route to early access, and it's a bad basis for anything that affects a team.
Patience is cheap here.
If you're preparing, the useful work is planning a short evaluation rather than guessing at settings. When Sonnet 5.5 does arrive, re-run the same handful of tasks you use for daily coding against it, instead of copying Opus 5.5 settings wholesale. Prompting habits changed between generations, so a fresh check beats an inherited config.
Then watch Anthropic's official channels for the model card and the price. The moment there's a published model ID and dollars per million tokens, the leak stops mattering and the real question starts: whether the new mid-tier earns a place in your workflow.
Until that happens, keep the assumption simple. The identifier exists in a third-party tool, developers report early access, and Anthropic hasn't said a word. Reported, unconfirmed, and worth watching rather than acting on.
