Claude Opus 5.5

0:00 Claude Opus 5.5 is the Claude model Anthropic tells most users to start with, released on Sep 22, 2026 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Anthropic says it performs "at the level of" Claude Fable 5.1 on most work, costs about 40% less to run than Opus 5 on typical workloads, and writes output more than 30% faster. It was the first model of the Claude 5.5 family, followed by Claude Sonnet 5.5 and Claude Haiku 5.5.

Quick facts

Sources: Models overview, Pricing and Model deprecations, Anthropic docs, read Nov 6, 2026; Plans and pricing, Claude, read Nov 6, 2026.

What changed from Opus 5?

0:58 The headline change is price. Every rate fell: input from $5 to $4, output from $25 to $20, cache reads from $0.50 to $0.20 and cache writes from $6.25 to $5 per million tokens. Combined with fewer tokens per task, Anthropic estimates about 40% lower cost than Opus 5 at default settings.

Other changes, from Anthropic's post and docs:

• Thinking can no longer be switched off. Adaptive thinking is always on, so the effort level is the main control (Effort, Anthropic docs, read Nov 6, 2026).
• Lower default effort. A request without an effort setting runs at medium, one level lower than on Opus 5.
• Stricter request rules. Forced tool use (tool_choice set to any or a named tool) and prefilled assistant turns return an error (Thinking, Anthropic docs, read Nov 6, 2026).
• Newer knowledge. The reliable knowledge cutoff moved from May 2026 to Jun 2026.

Claude Opus 5.5 benchmarks

Anthropic's launch table, at max effort unless noted, with production safeguards on:

Source: Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026. GPT-6 Astra leads on AutomationBench and Terminal-Bench-Science in Anthropic's own table. Anthropic gives a standard error of about 2.6 points on Terminal-Bench 4.0, and says some tasks were handled by fallback models because safeguards were on, which likely lowered scores. The Claude Sonnet 5.5 post a week later footnotes the 66.4% as an xhigh result, while the Opus 5.5 post says max effort unless noted; the two Anthropic pages disagree.

Independent results, as of Nov 6, 2026:

• Terminal-Bench 4.0: first on the public leaderboard at 64.85% (max effort, in Claude Code), ahead of Sonnet 5.5 at 61.82% and GPT-6.1 Sol at 58.18%. The run cost $4,718.65 against GPT-6.1 Sol's $634.24.
• Artificial Analysis Intelligence Index: 59 at max effort, second behind Gemini 4 Carbon (high) at 60; $10.04 per task at max and $3.46 at high.
• Arena: third in Text Arena at 1509, behind Gemini 4 Carbon and Gemini 4 Argon, and first in WebDev at 1810 (both tables dated Oct 27, 2026).

See Claude Benchmarks for how these tests differ.

Fast mode

Fast mode runs Opus 5.5 up to 2.5 times faster at twice the standard price: $8 input and $40 output per million tokens, charged across the whole context window. It works in Claude Code (toggle with Option+O on a Mac or Alt+O elsewhere), on the Claude API and on Claude Models on Google Cloud (Anthropic managed), but not with the Batch API (Interactive mode, Claude Code docs, read Nov 6, 2026). Opus 5 and Opus 4.8 fast mode costs $10 and $50, so the newer model is also cheaper here.

Where Claude Opus 5.5 is available

• Claude apps: Pro, Max, Ultra, Team and Enterprise; Free does not include Opus (Plans and pricing, read Nov 6, 2026). With the launch, Anthropic raised five-hour limits on Pro, Max, Team and seat-based Enterprise and gave subscribers a rate-limit reset they can save for later (Introducing Claude Opus 5.5).
• Claude Code: the default model on Pro, Max, Ultra, Team, Enterprise and the API (Model configuration, read Nov 6, 2026). Switch with /model opus.
• API and clouds: the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, with zero data retention available (Models overview).

Safety and known issues

Anthropic says Opus 5.5 got the best score of any model on its automated behavioral audit, tied Fable 5.1 for the lowest prompt injection success rate on a Gray Swan benchmark, and tried to get around its sandbox's limits about 85% less often than Opus 5 or Mythos 5.1 (Introducing Claude Opus 5.5). External testers included METR. Anthropic also names an open problem: Opus 5.5 often suspects it is being tested, which makes tests a weaker guide to real behavior. Most cybersecurity tasks such as exploit writing are rerouted to Opus 4.8, so security researchers outside the Cyber Verification Program will meet refusals or fallbacks.

What reviewers found

0:00 In a Sep 24, 2026 video, Dr. Károly Zsolnai-Fehér of Two Minute Papers had Opus 5.5 reproduce a research paper's honey-coiling simulation and said it ran at that quality in real time, which GPT-6 Astra had not managed in his tests. He then turned to the system card's risks: the model often suspects it is being evaluated, can run unattended for over 18 hours, and passed a strict quality bar in 16 of 18 tests, so he assumed it "can still hallucinate". Passing a test, he said, no longer necessarily means a model behaves well.

19:22 AI Search's Sep 23, 2026 review ran Opus 5.5 at max effort in Claude Code and the chat app. It solved a series of CAPTCHA-style puzzles, though one parking test took over ten minutes, which the host counted as a fail. In chat it circled a lizard that was not there in a hidden-frog image and named one of six brain tumor scans correctly, which the channel said tied the best result it had seen. The host called the model "incredibly expensive and I think overpriced" at about $6 per task on Artificial Analysis's index, while saying that for certain tasks it did seem to be the best model in the world.

Frequently asked questions

How much does Claude Opus 5.5 cost? Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on the Claude API, $2 and $10 through the Batch API, and $8 and $40 in fast mode. In the apps it is included in Pro and every higher plan.

Is Claude Opus 5.5 free? No. Claude's Free plan offers Sonnet 5.5 and Haiku 5.5; Claude Opus 5.5 needs Pro ($20 a month) or a higher plan, or API credit.

Is Claude Opus 5.5 better than Sonnet 5.5? In Anthropic's tables Claude Opus 5.5 is ahead on most tests, though Sonnet 5.5 reports a higher Terminal-Bench 4.0 score (70.6% against 66.4%). On the public Terminal-Bench leaderboard and Artificial Analysis's index, Opus 5.5 leads Sonnet 5.5 as of Nov 6, 2026, at twice the token price.

Can I turn off thinking on Claude Opus 5.5? No. Thinking is always on for Claude Opus 5.5; lower the effort level to use fewer tokens and get faster answers.

What is Claude Opus 5.5's context window? Claude Opus 5.5 has a 1M-token context window and writes up to 128K tokens per response, with the full window at the standard price.

Sources

• Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026
• Introducing Claude Sonnet 5.5, Anthropic, Sep 28, 2026
• Models overview, Anthropic docs, read Nov 6, 2026
• Pricing, Anthropic docs, read Nov 6, 2026
• Effort, Anthropic docs, read Nov 6, 2026
• Thinking, Anthropic docs, read Nov 6, 2026
• Model deprecations, Anthropic docs, read Nov 6, 2026
• Model configuration, Claude Code docs, read Nov 6, 2026
• Interactive mode, Claude Code docs, read Nov 6, 2026
• Plans and pricing, Claude, read Nov 6, 2026
• Terminal-Bench leaderboard, tbench.ai, read Nov 6, 2026
• Models leaderboard, Artificial Analysis, read Nov 6, 2026
• Text Arena and WebDev Arena, Arena, Oct 27, 2026
• Using Claude Opus 5.5 as your daily driver, Claude, Sep 23, 2026 (video)
• Claude Opus 5.5 AI: An Incredible Leap Forward, Two Minute Papers, Sep 24, 2026 (video)
• Claude Opus 5.5 is ridiculous, AI Search, Sep 23, 2026 (video)