Claude Opus 5.5: Anthropic's Flagship Model, Fully Specified

Claude Opus 5.5 is the vendor's current flagship model, released on September 22, 2026, and it sits in the middle of the Claude line as the vendor now documents it: cheaper and quicker than Claude Fable 5.1 above it, dearer and slower than Claude Sonnet 5.5 below it, and the direct successor to Claude Opus 5 at the top of the Opus line. The vendor's own model picker gives the shortest useful summary of where it belongs: start with Claude Opus 5.5 "for most workloads", and reach for Claude Fable 5.1 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." This page is the reference for the model that guidance points at — the envelope, the full rate card, the effort setting behind every published score, the independent read, and where it actually runs. Where a reader wants depth on one of those, this page routes rather than repeats: the rate-card arithmetic, the full benchmark sheet with its cross-model columns, and the migration details each live on their own page.

What Claude Opus 5.5 is, and where it sits in the line

Anthro​pic describes Claude Opus 5.5 as built "for long-running agentic coding and knowledge work", and introduces it as "the first model in our new Claude 5.5 family". It replaces Claude Opus 5 as the company's generally available flagship; Claude Opus 5 stays listed as a legacy model and is still served. The four-model line, as Anthro​pic's own models overview gives it, is short enough to hold in one pass:

• Claude Fable 5.1 — $10 input and $50 output per million tokens, comparative latency listed as slower, adaptive thinking always on, default effort high, 1M-token context, 128K max output, reliable knowledge cutoff June 2026. Anthro​pic positions it above the flagship and prices it accordingly.

• Claude Opus 5.5 — $4 input and $20 output per million, comparative latency moderate, adaptive thinking always on, default effort medium, 1M-token context, 128K max output, knowledge cutoff June 2026. It is the only model in the lineup that supports the effort parameter and does not default to high, and that detail matters more than it looks — see the effort section below.

• Claude Sonnet 5.5 — $2 input and $10 output per million, comparative latency fast, adaptive thinking, default effort high, 1M-token context, 128K max output, reliable knowledge cutoff June 2026, retirement no sooner than September 28, 2027.

• Claude Haiku 4.5 — $1 input and $5 output per million, comparative latency fastest, extended thinking, no effort parameter, 200K-token context, 64K max output, retirement no sooner than October 15, 2026.

Two readings of that table go wrong often enough to name. The first is the direction of the price move: Claude Opus 5.5 is cheaper than the model it replaced, not more expensive — $4/$20 against Claude Opus 5's $5/$25 — and it is cheaper than Claude Fable 5.1 by a factor of two and a half on input. The second is what the gap to Claude Fable 5.1 means. Anthro​pic reports Claude Opus 5.5 meeting or beating Fable 5.1 on several published tasks at a fifth to a third of the cost per task, while still directing genuinely demanding reasoning work upward; that is a cost-per-task claim about specific charts, not a general claim that the cheaper model is the stronger one, and the benchmark page keeps the distinction visible. Anthro​pic also says what comes next: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks", with "many of the same improvements to performance, efficiency, and safety". That sentence has since partly resolved: Claude Sonnet 5.5 shipped on September 28, 2026, while Claude Haiku 5.5 has not appeared — Anthro​pic's models overview still lists Claude Haiku 4.5 as the current small model, checked on October 7, 2026.

Released September 22, 2026 — and why the date is not the story here

The release date is documented twice on Anthro​pic's own surfaces: the announcement page is dated September 22, 2026, and the model page carries the line "Latest. Released September 22, 2026." The specification figures in this piece were read on September 28, 2026, from Anthro​pic's own documentation, from Artificial Analysis, or from our own model API, and each is labelled with which of those it came from; the statements in the availability section about what is current today were re-checked against those same surfaces on October 7, 2026, and are dated there.

Anthro​pic commits to supporting Claude Opus 5.5 "not sooner than September 22, 2027" on the platforms it operates itself; Amazon Bedrock and Goo​gle Cloud set their own lifecycle dates. Every Claude model ID is a pinned snapshot, including the dateless IDs used from the 4.6 generation onward, so claude-opus-5-5 is not a pointer that will drift under you.

Why this page is a reference, and not a second launch story

This is worth saying plainly, because the obvious reading of a page about a model released on September 22, 2026 is that it is launch coverage, and it is not. This is the canonical reference page for a model that is live in our catalogue today — the page a reader reaches after searching the model's own name rather than after reading a news feed. Its warrant is standing search demand that we measured first-party, not the freshness of a launch, and the September 22 release date appears here as a fact that dates the model, not as an event this piece is reporting.

That distinction is load-bearing. There are already launch posts, a leak write-up, comparison pages and several applied pieces about Claude Opus 5.5 in this blog's archive, and one more announcement would be a duplicate of work that exists. What did not exist until now was the page underneath all of them: the orientation a reader needs before any of the specialised pages is useful. So the pieces that do own a subject — pricing, benchmarks, and the API surface — are linked from here rather than re-derived, and this page sticks to its own job.

The envelope

From Anthro​pic's Claude Opus 5.5 model page and its models overview, both read on September 28, 2026:

• Context window — 1,000,000 tokens. Anthro​pic publishes no word count for the window, only a rough conversion for English text: "1 token is approximately 4 characters or 0.75 words," with the exact count varying by language and content type. One tokenizer caveat belongs beside that number: Claude 4.7 and later models, Claude Opus 5.5 included, use a newer tokenizer that produces approximately 30 percent more tokens for the same input, so a prompt sized against an earlier Claude model will not fill the same window.

• Maximum output — 128,000 tokens synchronously on the Messages API. On the Message Batches API it rises to 300,000 tokens behind the output-300k-2026-03-24 beta header.

• Modalities — Anthro​pic documents input as text and images, output as text, and lists vision, tool use and multilingual capability across the current lineup. Our own catalogue entry for the model additionally accepts a file input type, which the vendor page does not enumerate.

• Knowledge cutoffs — reliable knowledge cutoff June 2026; training data cutoff June 2026 as well.

• Model IDs — claude-opus-5-5 on the Claude API, on Goo​gle Cloud, on Microsoft Foundry and on Claude Platform on AWS; anthropic.claude-opus-5-5 on Amazon Bedrock.

• Status — Active (latest), with a retirement commitment of not sooner than September 22, 2027 on Anthro​pic-operated platforms.

The rate card, in full

All of the following is Anthro​pic's published pricing, read on September 28, 2026. Prices are per million tokens.

• Input — $4.00

• Output — $20.00. Thinking tokens bill as output tokens, and Claude Opus 5.5's adaptive thinking cannot be turned off, so thinking is not a separate line you can suppress.

• Cache read — $0.20, which Anthro​pic documents as 0.05x the base input price. That multiplier is a carve-out for this model: the standard rate on the sheet is 0.1x, and only Claude Fable 5.1 and Claude Mythos 5.1 sit lower at 0.025x.

• 5-minute cache write — $5.00, the standard 1.25x base input multiplier.

• 1-hour cache write — $8.00, the standard 2x multiplier. The minimum cacheable prompt length is 512 tokens.

• Batch API — 50% off both directions: $2.00 input and $10.00 output.

• Fast mode — $8.00 input and $40.00 output, exactly 2x standard, described by Anthro​pic as a research preview. It runs on the first-party Claude API only, is not offered on Claude Platform on AWS or the partner clouds, and is not available with the Batch API.

• US-only inference — setting inference_geo to "us" applies a 1.1x multiplier to input, output, cache writes and cache reads. Global routing, the default, uses standard pricing.

• No long-context surcharge — Anthro​pic states that Claude 4.6 and later models include the full 1M-token window at standard pricing, so a 900,000-token request bills at the same per-token rate as a 9,000-token one.

• Tool-use system prompt — Anthro​pic's own token-count table puts Claude Opus 5.5's tool-use system prompt at 286 tokens when tool choice is auto or none.

The cost question a reader actually has is not the rate card but the bill, and that is a different page's subject: the pricing page works through cost per finished task, what the cache line does across a long agent run, and how far the vendor's "about 40% less to run" characterisation can be verified from the sheet.

Effort, thinking, and the setting behind every published score

This is the part of the specification most launch coverage skips, and it is the part that changes what a benchmark number means. Claude Opus 5.5's adaptive thinking is always on and cannot be disabled — Anthro​pic says requests that set thinking: {"type": "disabled"} return a 400 error at every effort level, not just the high ones. Extended thinking, the manual budget_tokens mode, is the earlier mechanism: Anthro​pic documents it as deprecated on Claude Opus 4.6 and Claude Sonnet 4.6 and not accepted on later models. On Claude Opus 5.5 the control is therefore output_config.effort, and nothing else.

Five levels exist: low, medium, high, xhigh and max. Claude Opus 5.5 supports all five. Its default is medium — the only model that supports the parameter and does not default to high — which means a request migrated from Claude Opus 5 without an explicit effort value runs one level lower than it used to, and Anthro​pic's guidance is to run a fresh effort sweep on your own evals rather than carry settings across. Setting effort to the model's default produces exactly the same behaviour as omitting the parameter. Effort steers all tokens in the response, thinking included, and Anthro​pic is explicit that it is "a behavioural signal, not a strict token budget". Mid-conversation changes are supported through a per-message effort beta header, mid-conversation-output-config-2026-07-01, which keeps the prompt cache intact — a top-level change between requests does not.

The reason effort belongs on a reference page is that Anthro​pic's published results are not all run at one setting. The announcement carries a standing note — "Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort" — and then notes otherwise, per benchmark. Read that way, the sheet looks like this:

• Terminal-Bench 4.0 — 66.4%, reported at xhigh effort. This is one of the overrides, and it is a two-sided one: Anthro​pic reports GPT-6 Astra at high effort on the same chart and says both figures "represent each model's highest score". A comparison where one model is at xhigh and the other at high is not a matched-effort comparison, and Anthro​pic's own footnote is what tells you so.

• FrontierCode v1.1 (Main) — 54.6%, at default effort (medium). Anthro​pic notes it beats GPT-6 Astra's top score of 53.3% for about a fifth of the cost per task.

• CursorBench 4.0 — 52.5%, at default effort (medium). Against 51.8% for Claude Fable 5.1 at max effort and 46.6% for Claude Opus 5 at max effort. Note the asymmetry again: Claude Opus 5.5's number is produced one level below the other two models' best.

• GDPval-AA v2.1 — 1846 Elo, at max effort. Claude Fable 5.1 scores 1735 and Claude Opus 5 scores 1708 on the same chart.

• Everything else published — at max effort, by the standing note. That covers the OSWorld 2.0 and Chartography figures, whose labels ("partial" and "with tools") describe how the task was scored rather than which effort level produced it; neither chart carries an override, so both fall under the global sentence rather than escaping it.

Two further caveats belong beside those numbers rather than in a footnote. Anthro​pic publishes standard errors — ±2.6 points for Claude Opus 5.5 on Terminal-Bench 4.0, ±1.6 to 2 points for the other Claude models, and ±3.5 to 5 points per model on Terminal-Bench-Science 0.1 — and it checks two results against public leaderboards, reproducing Claude Opus 5's published 51.8% at 52.3% and its published 30.0% at 29.0%. Separately, its AutomationBench figures were produced without fallback models, which is a different harness condition from a deployment that has them. Anyone reading across models needs those three facts in hand. Claude Fable 5.1's independent standing is likewise measured on a different index revision from Claude Opus 5.5's, so the two are not directly comparable, and the benchmark page is where that gets worked through in the open.

The independent read

Anthro​pic's numbers are vendor-reported, and the sheet above is quoted as such. Independent measurement of Claude Opus 5.5 exists, and it comes from one place: Artificial Analysis, whose live model page we read on September 28, 2026.

Artificial Analysis scores Claude Opus 5.5 at 58 on the Artificial Analysis Intelligence Index, revision v4.3.2, ranking it #1 of 211 models on that revision. The configuration it is measured in is named on the page — "Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)" — and that label is the important part: the independent figure is a max-effort number, with fallback to the default configuration, so it is not directly comparable to any number in the vendor sheet that was produced at medium. Alongside the index, the same page reports output speed of 95.5 tokens per second, a cost of $5.98 to run the Intelligence Index per model, verbosity of 260 million output tokens, a 95% cache discount, a 1M-token context window, and a blended price of $2.94 per million tokens on a 7:2:1 cache-hit/input/output mix. Artificial Analysis publishes no Fable 5.1 variant alongside it, and no second Claude Opus 5.5 configuration, so the two models have not been benchmarked head to head there at matched effort and matched harness. That is a gap in the evidence, not a result, and it is why this page does not print an Opus-5.5-versus-Fable-5.1 comparison of its own.

One discrepancy is worth recording because it will resolve itself and a reader may see both versions. Our own model API, whose cached Artificial Analysis figure is dated September 22, 2026, carries 57.6 for the same index. Artificial Analysis's live page reads 58 on the current revision. The live page is the figure of record; the 57.6 is a snapshot taken on release day, on the revision in force then. Where this page states a single number, it is 58.

Our own telemetry is a third and different kind of measurement, and it is not a benchmark. Over the seven days ending September 28, 2026, requests routed to anthropic/claude-opus-5.5 through our playground showed a p50 time to first token of 4.69 seconds, a p95 of 10.00 seconds, 97.8 output tokens per second and a 5.59% error rate, with 62.7 million tokens routed in that window. Those describe our serving behaviour, not the model's capability, and they are quoted here so that no one reads them as a benchmark result.

Availability, and the version test

Anthro​pic says Claude Opus 5.5 is "now available on all platforms, including Amazon Web Services, Goo​gle Cloud, and Microsoft Azure", and the model page lists Claude API, Amazon Bedrock, Goo​gle Cloud, Microsoft Foundry and Claude Platform on AWS. Claude Opus 5.5 is also live on OrcaRouter as anthropic/claude-opus-5.5, at Anthro​pic's list price with no markup on top — $4.00 input, $20.00 output, $0.20 cache read and $5.00 cache write, the four figures read from our own model endpoint on September 28, 2026, matching the vendor sheet line for line. Our page for it carries the same 1M-token window and 128K maximum output, lists text, image and file as accepted input types and text as the only output, and marks the model as released 2026-09-22, featured, and not deprecated.

Everything else in the same catalogue is reachable from the same key: one API for more than 200 models, provider list price passed through at 0% markup, automatic failover when an endpoint degrades, and a routing DSL for the cases where you want to pin a model or blend several into one answer. For a version test that is the useful property. Because the markup is zero, the price you compare against is the same price Anthro​pic publishes, and because both the OpenAI-compatible chat-completions endpoint and Anthro​pic's own messages endpoint are served, an existing Claude client can be pointed at the new ID by changing two values — the model string and, if you want a different reasoning depth, the effort level. Run the same eval at medium and at xhigh, compare cost per completed task rather than cost per request, and let your own numbers decide whether the flagship or the model above it is the right default for the work in front of you. The API guide covers the parts that will not fail loudly when you swap the string; four breaking changes apply to code already running against Claude Opus 5, and one further change alters the response shape without raising an error at all.

Claude Opus 5.5 is still the top of the Claude line, and the date on that statement is October 7, 2026. Nothing above it in the family has been published: Anthro​pic's models overview lists exactly four current models — Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5 — with no fifth entry, the documentation page for Claude Opus 5.5 still carries the word Latest beside its name, and there is no successor identifier in the vendor's sitemap, in its deprecation table, in this catalogue or on Artificial Analysis. The same is true from the other direction: the ranked model list at Artificial Analysis, read the same day, still places Claude Opus 5.5 first. A reader who arrives here looking for the next Claude flagship is looking for something that has no name yet, and the practical answer is the one this section already gives: Claude Opus 5.5 is what you can call today at the vendor's $4 input and $20 output per million tokens, and Claude Fable 5.1 remains the step above it for the workloads where Anthro​pic recommends it.

One more thing belongs here, because a reader searching for a Claude model that does not exist yet meets the version numbers before the names. A 5.5 suffix does not mean the top of the line; it means the generation. Claude Sonnet 5.5 is the real 5.5 that shipped: released on September 28, 2026, vendor-priced at $2 input and $10 output per million tokens, and measured by Artificial Analysis at 56.00 on the same Intelligence Index revision, v4.3.2, that ranks Claude Opus 5.5 first — so a higher-sounding number on a smaller tier is still a smaller tier. The string "Claude Fable 5.5", by contrast, is a rumour label rather than a version number. The claim behind it was made in two social posts on October 2, 2026, and the chain ends there: as of today there is no model id, no rate card, no context window, no release date and no benchmark for it in the vendor's documentation, in this catalogue or on Artificial Analysis. The released top of the Fable line is Claude Fable 5.1, and the write-up on this blog under the title Claude Fable 5.5 Beat GPT-6.1 Astra Before It Exists — and That Is the Interesting Part is where that label's sourcing chain is examined; the distinction this page needs is narrower — a suffix that marks a generation against a name that marks a claim.

The short version

Claude Opus 5.5 is the model to reach for by default in Anthro​pic's current lineup, and the three facts that decide whether it is the right one for a given job are these. It defaults to medium effort where the models around it default to high, so a migrated setting silently changes behaviour. Its published benchmark figures are produced at different effort levels, and the two most-quoted ones — Terminal-Bench 4.0 and CursorBench 4.0 — are not comparable to the numbers beside them without reading the notes. And its independent score, 58 on Artificial Analysis's v4.3.2 index at ranked #1 of 211, is a max-effort measurement, which makes it the ceiling of the model rather than the thing you will run.

Below those three, the practical picture is straightforward: a million-token window with no long-context surcharge, 128K of synchronous output, a rate card that undercuts the model it replaced, adaptive thinking that cannot be switched off, and a retirement commitment two years out. Everything specific — the arithmetic of what a workload costs, the full benchmark sheet with its cross-model columns and each model's independent standing, and the migration details for code already running on Claude Opus 5 — lives on the pricing, benchmark and API pages respectively, and those are the next three clicks from here.