Claude Sonnet 5.5: The Droid Registry Leak and What It Proves

On September 27, 2026, Lumina posted that Claude Sonnet 5.5 looks basically staged for release: the API-style slug claude-sonnet-5-5 is already present in Factory Droid 0.228.0's model registry, wired behind a feature flag, while Factory's public model list still stops at Claude Sonnet 5. The same thread says Sonnet 5.5 is in greyscale testing, with replies claiming a lucky subset of Claude Code users can already hit it.

That is stronger evidence than a random screenshot of a chat title, because client registries are where vendors park real endpoint names before marketing turns the switch. It is still not a release. Anthropic has not posted Sonnet 5.5 pricing, benchmarks, or a model card. explainx.ai has not decompiled Droid 0.228.0 ourselves. Treat this as staging signal, the same category as arena nicknames for Gemini 4, not as something you should put in production config tonight.

TL;DR

Where this sits in Anthropic's September stack

Claude Opus 5.5 went live September 22, 2026 at claude-opus-5-5, with launch benchmarks, rewritten communication style, and cache-read pricing that matters for agents. explainx.ai covered that in the Opus 5.5 launch post.

In the same announcement, Anthropic said 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 official floor. Everything since is how soon and what the API string will be.

Sonnet is the tier most developers actually burn tokens on. Claude Sonnet 5 has been the default workhorse since June 2026, with permanent pricing documented in explainx.ai's Sonnet 5 guide and cost comparisons such as Sonnet 5 versus GPT-5.6 Luna Max. Opus 5.5 moved the ceiling; Sonnet 5.5 is the bet that Anthropic refreshes the middle of the lineup without waiting for a slow holiday cycle.

If Sonnet 5.5 lands with the same shape as Opus 5.5, the interesting number will not be a leaderboard screenshot. It will be dollars per completed task on Claude Code with your usual cache hit rate. explainx.ai already walked through that math for Opus in what a Claude Code task costs on Opus 5.5. Sonnet 5.5 will deserve the same treatment the day pricing exists.

What Lumina actually found

Lumina's claim has three separable parts:

1. Version: Factory Droid 0.228.0 (a specific client build, not Anthropic's server).
2. Artifact: an internal model registry mapping constant names to provider slugs.
3. Slug: claude-sonnet-5-5 sitting next to other Anthropic and OpenAI identifiers, including claude-opus-5-5, which is already a live public model ID.

The screenshot circulating with the post shows a minified registry fragment: assignment-style lines where human-readable keys map to quoted model strings. The claude-sonnet-5-5 entry is the one Lumina highlighted. The same frame also lists other slugs that may be placeholders, future SKUs, or unrelated experiments. Registry dumps are noisy. One credible Anthropic-shaped string is enough to pay attention; it is not proof that every string in the file will ship.

The second half of the leak is distribution, not naming:

• The slug is behind a feature flag in Droid, meaning the client knows how to request Sonnet 5.5 but most installs should not expose it in the UI yet.
• Factory's public model picker still shows Sonnet 5, which is what you would expect if Anthropic has not flipped the partner-facing switch.

That combination, hidden slug plus public list unchanged, is what people mean by staged for release. Engineering has wired the name; marketing and docs have not.

Greyscale testing and Claude Code

Lumina added that Sonnet 5.5 is in greyscale testing. In practice, that usually means Anthropic or a close partner routes a small percentage of requests to the new weights, or enables access for allow-listed accounts, while everyone else stays on Sonnet 5.

Replies on the thread repeat a familiar pattern from prior Claude rollouts: some Claude Code users see the new model if they are lucky. explainx.ai cannot verify who is in that cohort or whether the model string in their UI matches production weights versus a canary endpoint. If you are not already in a greyscale bucket, hunting for hidden flags in Droid is not a supported way to get early access, and it is a poor basis for team policy.

When Sonnet 5.5 does go wide, your prompts should not assume Opus 5.5 habits blindly. Anthropic's Opus 5.5 prompting guide documents effort defaults, communication changes, and tool-use patterns that differ from Sonnet 5 era habits. Plan to re-run a short eval on Sonnet 5.5 with the same tasks you use for daily coding, not to copy Opus settings wholesale.

What people are arguing about in the replies

"Model upgrades do not matter anymore." One reply argues Sonnet 5 already writes better code than most reviewers check, so the bottleneck is prompts, not weights. That is true for teams with weak review and vague tasks. It is less true for agents that run long loops, where marginal reliability and cost still move invoices. Sonnet 5.5 is aimed at the second group.

"You can always get cheaper though, this is a massive upgrade." Lumina's pushback matches Anthropic's Opus 5.5 story: the 5.x refresh was as much about efficiency and price per task as about raw IQ. If Sonnet 5.5 holds Opus-class coding closer to Sonnet pricing, it reopens the GPT-6 Sol versus Opus 5.5 conversation at the tier where most tokens actually live. That comparison already noted that a future Sonnet 5.5 may matter more for price-parity shopping than Opus 5.5 alone.

OpenAI DevDay timing. Several replies guess Anthropic ships Monday or Tuesday to step on OpenAI DevDay (September 30, 2026). That is narrative, not evidence. Vendors do time announcements for competitive air cover, but registry leaks have also appeared weeks before public launches. Do not schedule your launch around a dunk tweet.

Registry leaks versus official launches

explainx.ai has seen this movie with slug-first evidence:

The June 2026 Sonnet 5 rumor cycle followed a similar arc: leaks and codenames, then a formal launch with claude-sonnet-5. Sonnet 5.5 will probably rhyme with that pattern. The leak shortens the uncertainty window; it does not replace the announcement.

Third-party clients such as Factory Droid matter because they embed vendor model catalogs for power users who route Anthropic, OpenAI, and others from one harness. When Droid ships a registry entry, it often means someone on the partner side expects the endpoint to exist soon. It does not mean Factory speaks for Anthropic. For harness comparisons, Droid sits in explainx.ai's agent harness tier lists as a serious coding surface, which is why a Droid leak gets traction beyond Lumina's follower count.

What to do before Sonnet 5.5 is real

1. Pin your current default in Claude Code and any Droid config to a documented ID (claude-sonnet-5 or claude-opus-5-5). Do not point production at claude-sonnet-5-5 because a screenshot says so.
2. Save ten tasks you run weekly: a refactor, a test fix, a doc pass, a small feature, an agent loop with tools. You will rerun them on Sonnet 5.5 day one.
3. Read Opus 5.5's pricing move as a preview. If Sonnet 5.5 cuts output cost or improves cache behavior similarly, your loop and approval habits matter as much as the model swap.
4. Ignore DevDay bingo. If Anthropic ships during OpenAI's week, compare your eval, not the quote tweet ratio.

Honest limits

• explainx.ai did not verify Droid 0.228.0 artifacts locally.
• Haiku 5.5 was promised alongside Sonnet 5.5 but did not appear in this leak.
• Other strings in the same registry frame may be non-shipping placeholders.
• Greyscale quality can differ from GA weights; early wow posts are not benchmarks.
• Lumina is a leak account, not Anthropic; treat screenshots as unverified primary sources until docs land.

Bottom line

claude-sonnet-5-5 in Factory Droid 0.228.0 is credible staging evidence that Anthropic's next daily-driver refresh is close, aligned with coming weeks at Opus 5.5 launch and with greyscale whispers in Claude Code. It is not a release. Wait for Anthropic's model ID in official API documentation, pricing on the Claude platform page, and your own ten-task rerun. Until then, Sonnet 5 remains the supported workhorse, and Opus 5.5 remains the documented frontier for coding agents that need it.

Related reading

• Claude Opus 5.5 launch benchmarks and pricing
• Opus 5.5 prompting guide
• GPT-6 Sol vs Claude Opus 5.5
• Claude Sonnet 5 launch guide
• Sonnet 5 vs Luna Max on cost
• Opus 5.5 task cost on Claude Code
• Loop engineering for coding agents
• Agent harness tier list

Registry and greyscale claims reflect Lumina's September 27, 2026 post and public replies. Anthropic may ship different model IDs, pricing, or timing than leaks suggest.