Claude Opus 5.5

Spec

Attribute	Value
Developer	Anthropic
Released	2026-09-22
Announced	2026-09-22
Context window	1M tokens
Pricing	$4/M input · $20/M output · cache reads $0.20/M · cache writes $5/M
License	proprietary (API-only; no weight release)
Availability	Claude apps, Claude Code, Claude API as claude-opus-5-5, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry
Catalogue id	unknown
Rows this schema has no slot for, recorded here rather than invented as rows
(source):
max output 128K tokens (300k on the Message Batches API with the
output-300k-2026-03-24 beta header), adaptive thinking always on with a
default effort of medium, **training and reliable-knowledge cutoff Jun
2026**, and a retirement commitment of not sooner than 2027-09-22. A
fast mode is priced separately at $8/M input · $40/M output.

Catalogue id is unknown rather than absent: scripts/spec-check.py matches on the slug exactly, and whether OpenRouter serves this model under a name the slug claude-opus-5-5 reaches has not been checked from this sandbox, which is blocked from openrouter.ai. The daily spec-check Action is what will answer it.

Release Date

2026-09-22, available on all platforms the same day. Announcement and docs agree on the date.

Benchmarks

• 2026-10-25 external snapshots: Arena Agent / Overall lists Claude Opus 5.5 (Max) at 20.77% Net Improvement, 95% CI 18.39%–23.15%, in data published 2026-10-23. This is a signed contrast against a randomized baseline, not an Elo or success rate. (source)
• Artificial Analysis lists Claude Opus 5.5 (max with fallback) at 58 Artificial Analysis Intelligence Index and $5.98 Cost per Task USD. The latter is a measured task cost, not a token rate; the fallback configuration differs from Arena's label. (source)
• The October formal-verification experiment and its proof-budget conditions are recorded in Formal software verification. (source)

Anthropic's launch table, with the two comparison columns it published (source):

Benchmark	Opus 5.5	Fable 5.1	Opus 5
Terminal-Bench 4.0	66.4%	55.8%	52.3%
FrontierCode v1.1	54.4%	50.3%	48.0%
CursorBench 4.0	57.8%	51.8%	46.6%
GDPval-AA v2.1	1846 Elo	1735	1708
AutomationBench	40.0%	31.4%	26.9%
Humanity's Last Exam	67.7%	65.6%	63.6%
Terminal-Bench-Science 0.1	58.7%	52.6%	29.0%
OSWorld 2.0	81.8% partial	80.7%	74.0%
Chartography	89.0% with tools	88.4%	83.4%
Three cautions the table itself supplies. **No SWE-bench-family result is
headlined** — the coding claim rests entirely on Terminal-Bench 4.0,
FrontierCode v1.1 and CursorBench 4.0, so it cannot be compared against the
SWE-bench and DeepSWE numbers that most other pages on this wiki carry, and
specifically not against GPT-6 Sol's DeepSWE v1.1 figure published
the same day. OSWorld 2.0 is marked "partial" and Chartography "with tools";
both qualifiers are Anthropic's own and both are kept here. **No effort level is
stated for any row**, which matters because this model's default effort is
medium and Fable 5.1's is high — the comparison column may not be at the
same setting as the subject column, and nothing read says either way.

The Terminal-Bench-Science 0.1 row is the largest gap in the table: 58.7% against Opus 5's 29.0%, slightly more than double.

Use Cases

Anthropic names long-running agentic coding and knowledge work as the target. The docs make this concrete as a routing rule: start here for most workloads, and move to Claude Fable 5.1 only when evals on Opus 5.5 at higher effort still fall short (source).

The cost claim attached to the use case is 40% lower on real workloads, and Anthropic's stated mechanism is not only the per-token price: it says Opus 5.5 completes the same tasks with fewer tokens and generates output more than 30% faster. That is a claim about token economy, and nothing read this run measures it independently.

Preserved thinking is the release's other functional change. Anthropic describes it as an anti-distillation safeguard that "stops API users from editing Claude's prior context in an attempt to extract Claude's reasoning". It is a capability restriction shipped as a headline feature, which is the same move as the gated tiers on Claude Fable 5 and Open-Weights Policy Fight debates elsewhere on this wiki, applied to reasoning traces rather than to weights.

Safety gating: biology and cybersecurity safeguards fall back to Claude Opus 4.8 for restricted tasks, and full access requires verification programs. The model is available with zero data retention and carries watermarking measures per the EU AI Act.

2026-09-28 — a cheaper model from the same lab took this model's headline row

Claude Sonnet 5.5, released six days later at $2/$10 against this model's $4/$20, scores 70.6% on Terminal-Bench 4.0 where this model scores 66.4% (source).

That row matters more than a single benchmark normally would, because this model's launch table contains no SWE-bench-family result at all — the coding claim rests on Terminal-Bench 4.0, FrontierCode v1.1 and CursorBench 4.0. On the first of those three it is now second within its own vendor's line, to a model at half the price.

This model still leads on the other seven rows the two share, and the margins are mostly thin: GDPval-AA v2.1 1846 vs 1844, AA-Briefcase v1.1 1822 vs 1811, OSWorld 2.1 81.8% vs 80.1%, CursorBench 4.0 57.8% vs 55.5%, Chartography 64.4% vs 61.6%, Humanity's Last Exam 67.7% vs 64.5%, FrontierCode 1.1 Main 54.4% vs 46.2%. Six of the seven are under 4 points.

The comparison is not clean, and the reason is documented by the vendor. Neither launch table names an effort level for any row, and the two models default to different efforts — medium here, high on Sonnet 5.5 (docs). See Eval Harness Configuration. Anthropic's docs continue to recommend this model as the default for most workloads, which points the other way from the Terminal-Bench row.

Compared To

• Claude Fable 5.1 — the stated performance peer. Opus 5.5 beats it on all nine published rows while costing $4/$20 against $10/$50, a 2.5× gap on both input and output. Anthropic still routes demanding reasoning and long-horizon agentic work to Fable 5.1, so the table and the guidance point in different directions; the guidance is the more specific claim and is recorded as such.
• Claude Opus 5 — the predecessor, now listed by Anthropic under "Legacy models (still available)". 20% cheaper on input and output, 60% cheaper on cache reads ($0.20/M against $0.50/M).
• GPT-6 Sol and GPT-6 Luna — released by OpenAI the same day. Sol is $2/$10 against Opus 5.5's $4/$20, and claims roughly 80% lower cost per task than Fable 5 on DeepSWE v1.1. The two launches share no benchmark, so no direct comparison is available from vendor material; see Eval Harness Configuration.

Sources

• Introducing Claude Opus 5.5 (snapshot)
• Claude models overview — context window, max output, model IDs, cutoffs, retirement date

Conflicting Reports

Two Anthropic launch pages publish different Chartography and OSWorld figures for this model, and both are cited. Recorded here rather than resolved on the page, per the conflict rule in CLAUDE.md.

Claim	This model's own launch (2026-09-22)	Claude Sonnet 5.5's launch (2026-09-28)
Chartography	89.0% with tools	64.4%, in a row the table labels no tools
OSWorld	81.8% partial, named OSWorld 2.0	81.8% partial, named OSWorld 2.1
**The Chartography pair is almost certainly not a disagreement — it is the tool
condition**, which each table states and which the two tables do not share: 89.0%
is measured with tools and 64.4% in a no-tools row. Nothing read says so in as
many words, so it is not asserted as resolved; but a 24.6-point gap that lines up
exactly with a labelled condition is what a condition looks like, not what an
error looks like. The ## Benchmarks section keeps 89.0% with tools, being
this model's own launch figure with its own label
(source),
(source).

The OSWorld pair is stranger and smaller: the same value, the same "partial" qualifier, and a different version number six days apart. One of the two labels is wrong, or 2.1 left this model's score unchanged. Nothing read distinguishes them, and no figure on this page moves either way.

Both entries belong to Eval Harness Configuration's subject rather than to this model's capability: they were surfaced by claim-check.py, which compares a page's table against the snapshots it cites, and they are the kind of thing it exists to catch — a benchmark name that is stable while the condition under it is not.