Anthropic ships Claude Opus 5.5, and the headline is the price

Anthropic released Claude Opus 5.5 on Tuesday, and the most consequential number in the announcement is not a benchmark. The model lists at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Claude Opus 5 — a 20% cut at the top of the company's widely available lineup, while the documentation now recommends Opus 5.5 rather than any other model as the default starting point for most workloads.

The API id is `claude-opus-5-5`. The context window stays at 1M tokens, max output at 128K, and the training data cutoff at June 2026. It is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS from day one.

Where the 40% actually comes from

Anthropic's framing is that Opus 5.5 "costs 40% less to run than Opus 5" on typical workloads. That is a workload claim, not a price cut, and the two numbers are worth separating.

• Per-token list prices fell 20%, not 40%: $5 to $4 on input, $25 to $20 on output.
• Cache reads fell further — $0.20 per million tokens, which Anthropic puts at 60% below Opus 5. The docs explain why: cache reads are normally 10% of the base input price, but 5% on Opus 5.5.
• Output generation is more than 30% faster, so the same job occupies less wall-clock time.
• The default effort level moved from `high` on Opus 5 to `medium` on Opus 5.5. A request that omits `effort` now thinks less by default, and therefore bills fewer tokens.

That last point is not a footnote. Anthropic's own migration guide lists "re-run your effort sweep" as its first recommended change, noting that effort is now the only thinking control and that its default has dropped a level. Teams that upgrade by swapping the model string and changing nothing else will see a cheaper bill partly because the model is cheaper and partly because it is, by default, doing less work per request. Whether that trade holds is exactly what the effort sweep is for.

The benchmark claim: beating the bigger model

The more striking assertion is that a cheaper Opus now matches or beats Claude Fable 5.1 — Anthropic's most capable widely released model, which lists at $10 and $50 per million tokens — on a good deal of work. On the company's published figures, comparing Opus 5.5, Fable 5.1 and Opus 5:

• Terminal-Bench 4.0: 66.4% vs 55.8% vs 52.3%
• FrontierCode v1.1: 54.4% vs 50.3% vs 48.0%
• CursorBench 4.0: 57.8% vs 51.8% vs 46.6%
• OSWorld 2.0 (computer use): 81.8% vs 80.7% vs 74.0%
• GDPval-AA v2.1: 1846 vs 1735 vs 1708 Elo

These are Anthropic's own evaluations, reported by Anthropic, and no independent replication existed at publication time. The docs still steer demanding reasoning and long-horizon agentic work to Fable 5.1, with a revealing piece of guidance: use Fable 5.1 when "your evals on Claude Opus 5.5 at higher effort still fall short." The recommendation, in other words, is to try the cheaper model harder before paying for the larger one.

Four things that break on upgrade

This is not a drop-in swap for code running on Opus 5. The migration guide documents four breaking changes, each returning a 400 rather than degrading quietly:

1. Thinking cannot be disabled. Both `thinking: {"type": "disabled"}` and the older `budget_tokens` form are rejected. Effort is the only control left; where you previously disabled thinking to save tokens, you now pick a lower effort.
2. Forced tool use is gone. `tool_choice` types `any` and `tool` return a 400, including on the token-counting endpoint. The replacement is `auto` plus strict tool use or structured outputs, and saying in the prompt when the tool applies.
3. Thinking blocks are bound to the model and the conversation. Only Fable 5.1 and Mythos 5.1 can read Opus 5.5's thinking blocks, so a router or fallback that moves a conversation elsewhere runs those turns without them. Keep the conversation append-only.
4. Computer use moved to a new toolset. On the Claude API and Google Cloud, `computer_20251124` is rejected in favour of `computer_toolset_20260801`, which changes the agent loop, not just the request. On Amazon Bedrock the older tool keeps working.

There is a quieter change that will not fail a request but will visibly change a product: text the model writes between tool calls now comes back as progress-update `thinking` blocks rather than `text` blocks, and at the default display setting their content is empty. An application that streams that narration to users as progress goes silent between tool calls until it opts into `display: "updates"` or `"summarized"`.

Developers should also expect a wider set of refusals. Opus 5.5's safety classifiers cover more categories than Opus 5's, so `stop_details.category` values such as `bio` and `reasoning_extraction` join `cyber` — and server-side fallback does not retry requests declined under `reasoning_extraction`.

Released into an argument Anthropic started

The timing is the part worth sitting with. This is Anthropic's first model release since CEO Dario Amodei publicly argued that AI progress should be slowed so that safety work can keep up — a position the announcement does not walk back so much as reframe.

Pacing is an approach to keeping AI safe, remaining competitive with China, and realizing AI's benefits, particularly in areas like biology and medicine.

The safety claims attached to the release are specific. Anthropic says Opus 5.5 is the strongest-performing model it has tested on its automated behavioural audit, its most comprehensive alignment test, run across roughly 2,000 scenarios. The model was evaluated before release by outside organisations including METR and Frontier Design, and ships with safeguards matched to Fable 5.1's standards for cybersecurity and biology, plus anti-distillation protection. Access for legitimate high-risk work is routed through two gated channels: the Life Sciences Verification Program for biology research and the Cyber Verification Program for vetted security practitioners.

The honest tension is that pacing and a 40% cost reduction pull in different directions. Cheaper frontier-class inference means more of it, from more people, faster — which is the stated point of the release and also the thing pacing is meant to moderate. Anthropic's answer is that the safety practices scale with the capability, and that the gating programmes handle the sharp edges. It is a coherent position; it is not a slowdown, and the company does not claim it is one.

Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks, which would make this the first of a wave rather than a one-off. For anyone currently paying Opus 5 or Fable 5.1 rates for agentic coding, the immediate work is unglamorous and worth doing: run your own evals at more than one effort level before assuming either the savings or the quality.