Claude Opus 5.6: The Search Exists, the Model Does Not — Here's What Is Actually There

Claude Opus 5.5code>claude-opus-5-6/code> model ID, no rate card and no documentation page. The legacy list underneath it goes back through Claude Opus 5, Opus 4.8, 4.7, 4.6 and 4.5 — five older Opus builds, none of them a 5.6.

That gap between demand and documentation is the whole article, and it is worth writing down accurately rather than filling with speculation. People are searching for a model that has not shipped, and the search results they get are landing on pages about a different company's model.

Why "5.6" keeps appearing

There are three unrelated reasons the string turns up, and separating them explains most of the confusion.

It belongs to OpenAI. GPT-5.6 is a real and current family — Sol, Terra and Luna — with Sol generally available since 9 July 2026 at $5 input and $30 output per million tokens. Anthropic's own comparison material uses the name extensively, as do ours and everyone else's. A reader scanning a page about Claude Opus 5.5 versus GPT-5.6 Sol can easily come away with the impression that there are two 5.6s in the conversation. There is one.

The Opus line has moved fast enough to make the next number feel overdue. Opus 4.6, 4.7, 4.8, 5 and 5.5 arrived in sequence, and Opus 5.5 landed on 22 September 2026 with a price cut attached — $4 and $20 against Opus 5's $5 and $25. After a run like that, a 5.6 in October is a reasonable guess. It is still a guess.

The Fable rumour leaks sideways. Claude Fable 5.5 has been the subject of six weeks of checkpoints claimed by a single X account, including one "delivered to partners" on 18 October. Because Fable and Mythos are the frontier track and Opus is the volume track, chatter about one tends to produce searches for the other. Nothing about the Fable 5.5 claims implies an Opus 5.6.

The four surfaces a real model appears on

Every Claude launch this year has produced the same four artefacts together, and they are a cheap test to apply to any rumour.

• A row in Anthropic's models overview carrying a model ID, a rate card, a context window and a retirement commitment.

• A documentation page that resolves at its own path, rather than returning a 404.

• A line in the API release notes naming the model and its launch date.

• A row on the independent scoreboards, which lag but do not stay empty for a shipping model.

Today the first of those lists four models and the second is the page you would be reading. The failure is not partial: there is no fifth row, no orphan path, and no release-note entry after Claude Haiku 5.5 on 7 October — everything since has been platform work, fast mode leaving research preview for Claude Opus 5.5 on 15 October, a Compliance API beta the same day, OpenTelemetry export and Managed Agents scheduled deployments on 14 October.

An absence proves nothing about a company's internal roadmap. It does establish the frame: a model that would be callable today has four fingerprints, and none of them are present.

The precedent worth remembering

Six weeks ago the same shape appeared one tier up. On 15 October a well-followed account said it expected Claude Fable 5.5 "next Tuesday"; on 19 October the account posting Fable demos all month said the Blender render everyone had seen was Fable 5.5 at its first checkpoint and that a second checkpoint had been "delivered to partners yesterday" and was "better than first".

Every checkable surface disagreed. The models overview listed four current models, with no Fable 5.5 row. The Fable 5.5 documentation path returned a 404. The release notes stopped at Haiku 5.5. Artificial Analysis carried no Fable 5.5 entry. Then the Tuesday the rumour named arrived, and what arrived instead was a sentence about a checkpoint.

That is not a debunk — a team shipping checkpoints to partners is a team that is close, and partner delivery is the staging step that leaves fingerprints outside the lab. But it is a useful calibration: a claim with a date can be graded by the calendar, and this one was. "Opus 5.6" has not even produced a claim with a date.

If what you wanted was the current Opus

Claude Opus 5.5 is the answer, and it is a more interesting model than its position in the lineup suggests. Released 22 September 2026, it is the first Opus to undercut its own predecessor and the model Anthropic's own documentation tells most people to start with: 1M-token context, 128K maximum output, knowledge cutoff June 2026, retirement no sooner than 22 September 2027, cache reads at $0.20 per million, batch at half price.

Three things about it surprise people migrating from Opus 5. Thinking cannot be disabled at all — code>{"type": "disabled"}/code> and code>budget_tokens/code> both return a 400 at every effort level, so effort is the only control and its default is code>medium/code>, one step below Opus 5's code>high/code>. Forced tool use returns a 400, so code>tool_choice: any/code> and named-tool forcing are gone in favour of code>auto/code> plus code>strict: true/code> or structured outputs. And on the Claude API and Google Cloud, computer use only works through the code>computer_toolset_20260801/code> tool.

On the independent scoreboard it is the highest-scoring Claude model: 58 on the Artificial Analysis Intelligence Index at maximum effort with fallback, at $5.98 per completed index task, which puts it above Claude Fable 5.1's 53 while costing less per task. The same board lists it a second time at xhigh effort, scoring 56 at $3.46 — the configuration caveat that applies to every ranking.

What to do while you wait

Nothing in a production path should be planned around a model number that does not exist. What is worth doing is making the swap cheap for the day a fifth row appears in that table.

• Keep the model ID in configuration, not in code. Every Anthropic ID is a pinned snapshot, so a migration is a config change and a re-run of your evals, not a rewrite.

• Set effort explicitly. Opus 5.5's default is medium and its scores move two points between medium-adjacent and maximum settings; if you never set it, you are running an unmeasured configuration.

• Route rather than commit. Claude Opus 5.5 is in the OrcaRouter catalogue as code>anthropic/claude-opus-5.5/code> at Anthropic's list price with 0% markup — $4 and $20 with the $0.20 cache read — on the same key as Claude Fable 5.1 and Claude Sonnet 5.5. Putting a new model on a slice of traffic behind the same OpenAI-compatible endpoint, with a failure-rate or latency threshold sending requests back to the proven one, is what turns "should we move to 5.6" into an afternoon of evals instead of a migration.

What would change this page

A fifth row in the models overview with a model ID next to it. If that appears, this article is wrong by the afternoon and should be rewritten rather than amended — a released model deserves its own page, not a correction on a page about its absence.

Until then the accurate answer to "when is Claude Opus 5.6 coming out" is that Anthropic has not announced it, has no documentation for it, and lists four current models, none of which is it. The best available Opus is 5.5, it is cheaper than the Opus 5 it replaced, and on the one independent index that scores both, it outperforms the more expensive Fable tier. That is a less exciting sentence than a release date and a more useful one.

Sourcing note: the four-model lineup, model IDs, prices, context windows, retirement commitments and the migration restrictions are Anthropic's Claude Platform documentation, read on 26 October 2026. The GPT-5.6 Sol release date and price are OpenAI's launch material. The Fable 5.5 checkpoint claims are from the X posts described, dated 15 and 19 October 2026, and remain unverified. Intelligence Index and cost-per-task figures are Artificial Analysis's, read on the same date. Prices and route availability are OrcaRouter's catalogue and the public model list on our gateway.