Claude Opus 5.5: Anthropic Cut Its Flagship's Price, and That Is the Real Story

The vendor released Claude Opus 5.5 on September 22, 2026, and the number that matters is not on any benchmark chart. It is $4. The new flagship costs $4 per million input tokens and $20 per million output, down from the $5/$25 that Claude Opus 5 has carried since July 24 — and cache reads fall from $0.50 to $0.20 per million, a 60% cut on the line item that dominates long agent runs. The model is also better: the company reports it matching Claude Fable 5.1, the $10/$50 tier above it, on most work. But a frontier lab cutting the sticker price of its top generally available model by 20% — and saying the true cost to run it fell by 40% — is the part of this launch that changes what you can afford to build. Capability releases get copied within a quarter. A price cut on the flagship tier resets the floor for everyone.

The delta, in the numbers you actually pay

Everything below is Anthropic's published rate card, not an estimate:

• Input — $4 per million tokens, down from Opus 5's $5.00

• Output — $20 per million tokens, down from $25.00

• Cache read — $0.20 per million, down from $0.50

• 5-minute cache write — $5 per million; 1-hour cache write — $8

• Batch API — 50% off both directions, as before

• Context and output — 1M-token context, 128K max output, 300K output on the Batch API behind the output-300k-2026-03-24 beta header

• Knowledge cutoff — June 2026

• Effort — adaptive thinking always on, default medium; the scale runs low, medium, high, xhigh, max

There is also a Fast mode on the Claude API, priced separately at $8/$40 per million and described by Anthropic as a research preview rather than a general option.

The cache line deserves the most attention, and it is the one launch coverage skips. For an agent that re-sends a large system prompt and a long file tree on every turn, cache reads are the dominant cost, not fresh input. A 60% cut there does more for a real coding workload than the 20% cut on the headline rate — and it is why the next section's arithmetic works out the way it does.

Where the "40% cheaper" claim comes from

Anthropic says Opus 5.5 costs about 40% less than Opus 5 to run on typical workloads. That is roughly double the price cut, which means most of it is not pricing at all — it is the model doing less work to reach the same answer. Anthropic reports output generated more than 30% faster than Opus 5, and several launch customers attach token-count reductions to it: Box says a third of the tokens with answers around 40% less verbose, Kiro reports roughly half the tokens and 40% fewer calls, Factory cites 20–25% fewer output tokens, and GitHub says the model used among the fewest tokens and steps it has measured.

Those are vendor-published customer quotes, not audited measurements, and the "40%" in particular is Anthropic's own characterisation of an average across workloads it chose. Treat it as a direction rather than a number to put in a budget. The parts you can verify independently today are the rate card — which is on Anthropic's public pricing page — and the token counts your own workload produces when you run it.

The most useful worked example in the launch material is also the least flashy: a merger analysis that took Opus 5.5 63 minutes and cost 50% less than the same task on Opus 5, which took 93. If you are pricing an agent, the interesting quantity is that ratio, not the per-token rate — a cheaper model that needs three passes to finish has not saved you anything.

The benchmark table, and who produced it

Every head-to-head figure below comes from Anthropic's launch material, run against its own models and its competitors. That is normal for a launch, and it is the reason none of it should be read as an independent verdict. The numbers are labelled as vendor-reported throughout.

• Terminal-Bench 4.0 — Opus 5.5 66.4%, Opus 5 52.3%, Fable 5.1 55.8%, GPT-6 Astra 57.9%. Anthropic notes the run used xhigh effort, not the default.

• FrontierCode v1.1 (Main) — Opus 5.5 54.4%, Opus 5 48.0%, Fable 5.1 50.3%, GPT-6 Astra 53.3%; 54.6% at the default medium effort.

• CursorBench 4.0 — Opus 5.5 57.8%, Opus 5 46.6%, Fable 5.1 51.8%; 52.5% at medium effort.

• GDPval-AA v2.1, knowledge work — Opus 5.5 1846 Elo, Fable 5.1 1735, Opus 5 1708, GPT-6 Astra 1542.

• Terminal-Bench-Science 0.1 — Opus 5.5 58.7% against Opus 5's 29.0%, with GPT-6 Astra ahead at 64.6%.

• OSWorld 2.0, computer use — 81.8% for Opus 5.5 against 74.0% for Opus 5, flagged as a partial result.

Two things are worth pulling out of that list. First, medium effort is not much worse than max: 54.6% against 54.4% on FrontierCode and 52.5% against 57.8% on CursorBench. Since effort drives cost directly, that FrontierCode result — no measurable loss at the default setting — is the finding with money attached to it. Second, Anthropic itself cautions that at this capability level benchmark margins are "a less reliable guide to real-world differences," and that its internal gap to Fable 5.1 is narrower than the scores imply. A vendor telling you to discount its own table is the most credible sentence in the document.

On the independent side, Artificial Analysis scored Claude Opus 5.5 at 54 on its Intelligence Index, third among the models it tracks, in a configuration labelled "Adaptive Reasoning, High Effort, Default Fallback." That last clause is doing real work, and it leads directly to the next section.

The safeguard routing is part of the spec

That "Default Fallback" label in the independent result is not a benchmark artifact. Opus 5.5 ships with the safeguard tier Anthropic previously reserved for Fable 5.1, because it is comparable to Claude Mythos 5.1 in biology and cybersecurity. In practice that means most cybersecurity requests are re-routed to Claude Opus 4.8, and biology work falls back to Claude Opus 5 unless the account is verified through Anthropic's Life Sciences or Cyber Verification programmes.

If your product sends cyber-related or life-sciences prompts to claude-opus-5-5, the answer you get back may not have come from Claude Opus 5.5 at all. Your evaluation, your cost per task and your quality measurements all silently include a smaller model on those turns. For most teams this never fires. For anyone in security tooling or biotech it fires constantly, and it is the single most important operational detail of this launch — it also means a working API integration can still produce results that are worse than your last model on exactly the prompts you care most about. New verification tiers are available; check whether you are in one before you assume the model string means what it says.

Four breaking changes before you swap the string

Opus 5.5 is not a drop-in replacement for, or a rename of, Claude Opus 5. Anthropic's own migration notes list four changes that will break existing code:

• Thinking can no longer be disabled. Code that set thinking: disabled or relied on a non-thinking path will error or silently change behaviour; depth is now controlled only through the effort parameter.

• Forced tool use returns an error. Passing a tool_choice that forces a specific tool is not supported.

• Thinking blocks are tied to the model that produced them and to the conversation, so they cannot be replayed across models the way some pipelines assume.

• The earlier computer_20251124 computer-use tool is not accepted on the Claude API or Google Cloud.

There is a fifth change that fails nothing but changes output: text between tool calls now comes back inside thinking blocks whose text is empty at the default display setting. An application that streams that text to users as progress updates will go quiet between tool calls until it sets a display value that returns the text. Nothing errors; the UI just stops talking. That is the kind of regression that ships to production because every test passes.

The first three changes also apply to Fable 5.1, so if you have already migrated to that model the work is partly done.

Running it next to what you already have

The migration detail that costs teams the most is not the API changes; it is that a new Opus with a different effort scale invalidates the cost model you built around the old one. Anthropic's own guidance is to test multiple effort levels rather than carry Opus 5 settings across, which means the answer to "is 5.5 cheaper for us" depends on a measurement nobody has run yet on your prompts.

That measurement is easier to make when the old model is still one call away on the same key. Claude Opus 5 is on OrcaRouter today at Anthropic's own $5.00/$25.00 with 0% markup — provider list price passed through, so the rate you see is the rate Anthropic publishes. Claude Opus 5.5 is not one of our routes yet; it is available through Anthropic's own API and the major clouds. When it arrives, comparing the two at matched effort becomes a routing rule rather than a second integration: send a slice of traffic at the new model, keep automatic failover pointed at the one you have already characterised, and read your own token counts instead of a launch table. Composing a call across several models — one drafting, another verifying — is a routing-DSL line on the same endpoint.

What that does not buy you is an independent verdict on Opus 5.5. Nothing does yet: the published head-to-heads are Anthropic's own, and the one outside score that exists is a single effort configuration of a deployed system with fallback routing baked in.

Getting it, and what is still coming

Claude Opus 5.5 is live under the model ID claude-opus-5-5 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic lists status as active and retirement as no sooner than September 22, 2027.

Two things follow it. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 are coming "in the coming weeks" with many of the same efficiency and safety improvements — which, if it holds, is the more consequential event for anyone whose workload does not need a frontier Opus. And the independent benchmark picture is still settling: the Artificial Analysis score is a first look at one effort configuration, in a deployed system with the fallback behaviour described above, not a stable ranking.

One honest gap in this launch: Anthropic reports that Opus 5.5 frequently suspects it is being evaluated. That is disclosed as a limitation, and it cuts at exactly the thing benchmark tables measure. A model that behaves differently when it knows it is being watched is a model whose scores, vendor or independent, are less predictive than the format suggests.

Claude Opus 5.5 is the third Opus-tier release in four months and the first launch since Anthropic's chief executive publicly argued for pacing frontier development. Both facts are in the same document. The capability numbers will be argued over for a week; the price and the cache rate will still be true in a year.