更新
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic's September 2026 Opus model for agentic coding, long-running agents, and document and spreadsheet work. Released September 22, 2026, it is the first model of the Claude 5.5 family and replaces Claude Opus 5 in the Opus line. On UnificAlly it runs as anthropic/claude-opus-5-5. It takes text and image input, outputs text, and carries a 1M token context window with up to 128k output tokens. Adaptive thinking is always on. The headline is that an Opus-tier model now matches Claude Fable 5.1, Anthropic's top generally available model, on most of the evaluations Anthropic published, while generating output about 30% faster than Claude Opus 5.
What's new in Claude Opus 5.5
Agentic coding: 66.4% on Terminal-Bench 4.0
Claude Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra, 55.8% for Claude Fable 5.1, and 52.3% for Claude Opus 5. It also leads FrontierCode v1.1 at 54.4% and CursorBench 4.0 at 57.8%, both ahead of Fable 5.1 (50.3% and 51.8%).
#1 on the Artificial Analysis Intelligence Index
The independent Artificial Analysis index scores Claude Opus 5.5 at 58 at max effort, first of 212 models, ahead of GPT-6 Astra and Claude Fable 5.1 at 53, as of September 23, 2026. Anthropic also reports output about 30% faster than Claude Opus 5.
Knowledge work and computer use
On GDPval-AA v2.1, real occupational tasks scored by Elo, Claude Opus 5.5 reaches 1846 against 1735 for Claude Fable 5.1 and 1708 for Claude Opus 5. Humanity's Last Exam lands at 67.7%, OSWorld 2.0 computer use at 81.8% (Opus 5: 74.0%), and AutomationBench business workflows at 40.0%, just behind GPT-6 Astra at 41.4%.
Fewer tokens for the same job
Early testers report the gain comes with less work, not more. GitHub saw about half the tokens and steps on terminal tasks, Factory measured 20 to 25% fewer output tokens than Opus 5, and Anthropic cites a 200,000-line codebase audit finished in under three hours where Opus 5 took more than 20. Effort now defaults to medium, so routine calls stay short unless you ask for depth.
What Claude Opus 5.5 is best for
Agentic coding
Terminal-heavy agent loops, multi-file changes, and large migrations that run unattended.
Codebase audits
Reads and reasons across hundreds of thousands of lines inside the 1M token window.
Knowledge work
Financial models, legal research, and reports built from long source documents.
Computer use
Operates browsers and desktop apps, with 81.8% on OSWorld 2.0.
Charts and documents
Reads dense charts, tables, and scanned pages from image input.
High-volume agents
Fewer output tokens per task and a medium effort default keep long runs lean.
Claude Opus 5.5 use cases
Run it as the default model behind a coding agent: point it at a repository, a failing test suite, and a terminal, and let it iterate until the build is green. Use it for code migrations that used to need a week of supervision; Anthropic reports a 680,000-line migration finished in under a day and a C-to-Rust translation of HAProxy done in 9.5 hours. Put it behind finance and legal research products that read long filings and contracts with charts inline and return structured briefs. It also fits browser and desktop automation, where it now edges out Claude Fable 5.1 on OSWorld 2.0.
Claude Opus 5.5 limitations
Thinking can no longer be disabled. {type: "disabled"} returns an error at every effort level, so the only way to make calls cheaper and faster is to lower effort. Forced tool use also returns an error: keep tool_choice on auto, name the tool in the prompt, and use strict tool schemas when you need valid arguments. Thinking blocks are tied to the model and the conversation, so editing earlier turns invalidates them; keep conversation history append-only.
Safety classifiers are broader than on Claude Opus 5. Biology and reasoning-extraction checks now sit alongside the cybersecurity classifier, and Anthropic routes most cybersecurity tasks to Claude Opus 4.8. Handle the refusal stop reason, since some benign requests near those areas still get declined. GPT-6 Astra keeps a lead on AutomationBench (41.4% against 40.0%) and Terminal-Bench-Science (64.6% against 58.7%). Output is text only.
Claude Opus 5.5 vs Claude Fable 5.1
Same 1M token context, same 128k output limit, same always-on thinking. On the evaluations Anthropic published, Opus 5.5 comes out ahead everywhere: Terminal-Bench 4.0 at 66.4% against 55.8%, GDPval-AA v2.1 at 1846 against 1735, Humanity's Last Exam at 67.7% against 65.6%, Terminal-Bench-Science at 58.7% against 52.6%, OSWorld 2.0 at 81.8% against 80.7%, and Chartography at 89.0% against 88.4%. Artificial Analysis agrees, with 58 against 53 on its Intelligence Index. It is also faster. Arena.ai has no Opus 5.5 votes yet as of September 23, 2026, so crowd-voted rankings are still pending. Our pick: for new agentic coding and knowledge-work builds, start on Opus 5.5 and only reach for Fable 5.1 if a specific task measures better there.
When to use Claude Opus 5.5
Use Claude Opus 5.5 as the default Claude model for serious agentic and coding work, and as the upgrade path for anything running on Claude Opus 5 or Claude Opus 4.8. Set effort explicitly: high or xhigh for long agent runs and hard debugging, medium (the default) for everyday calls, low for sub-agents and simple extraction. For high-volume, latency-sensitive chat, Claude Sonnet 5 remains the lighter option.
Claude Opus 5.5 API examples
Call Claude Opus 5.5 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-opus-5-5.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "anthropic/claude-opus-5-5",
"messages": [
{ "role": "user", "content": "Here is our Express API. Migrate every route to Fastify, keep the tests passing, and list anything you could not port." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQ
Claude Opus 5.5に関するよくある質問
anthropic/claude-opus-5-5, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
Anthropic released Claude Opus 5.5 on September 22, 2026, as the first model of the Claude 5.5 family. It is live on UnificAlly since September 23, 2026.
A 1M token context window with up to 128k output tokens per response, the same limits and tokenizer as Claude Opus 5.
On Anthropic's own evaluations it matches or beats it. Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1, 1846 on GDPval-AA v2.1 against 1735, and 67.7% on Humanity's Last Exam against 65.6%. Artificial Analysis puts it #1 on its Intelligence Index at 58, ahead of Fable 5.1 at 53, as of September 23, 2026.
Adaptive thinking is always on and cannot be disabled. The effort setting from low to max controls how deep it goes, and the default is medium, one level below Claude Opus 5's default of high. Raise it for hard agentic work.
Four things. Thinking can no longer be turned off. Forced tool use (tool_choice of any or a named tool) returns an error, so use auto and name the tool in the prompt. Thinking blocks are tied to the model and the conversation, so keep history append-only. And the default effort drops to medium, so set it explicitly if your prompts were tuned for high.
Yes. It takes text and image input in one request, and scores 89.0% on Anthropic's Chartography chart-reading evaluation. Output is text.
Claude Opus 5.5のガイドと比較
このモデルを実行する前に、料金、制限、テスト済みの出力をご確認ください。
