Claude Opus 5.5: Benchmarks, Pricing, and What Changed

What’s New in Claude Opus 5.5

Opus 5.5 succeeds Claude Opus 5, which is now a legacy model (still available, retirement no sooner than July 24, 2027). Anthropic built it for long-running agentic coding and knowledge work, and the headline changes are about doing the same work with less:

• Cheaper per token: $4/$20 per million tokens versus Opus 5’s $5/$25 — 20% off input and output. Cache reads drop 60%, from $0.50 to $0.20.
• Fewer tokens per task: Anthropic says typical workloads cost about 40% less than on Opus 5 once token efficiency is counted. Customers report the same pattern: Box measured a third of the tokens Opus 5 used, and Optiver cut the cost of one workload by 40–50%.
• Faster: output generation is more than 30% faster than Opus 5, and a new fast mode (research preview) runs up to 2.5× faster at $8/$40 per million tokens.
• Clearer writing: less jargon, with the important information placed at the start of its responses.
• Newer knowledge: a June 2026 reliable knowledge cutoff, up from Opus 5’s May 2026.
• Default effort is now medium on the API (Opus 5 defaulted to high), and adaptive thinking is always on.

Early customer results in Anthropic’s announcement point the same way. Deloitte reported Opus 5.5 caught 72% of known bugs in a code-review test versus 56% for Opus 5, and Hebbia measured 86.6% of its target criteria versus 60.3% for Opus 5.

Claude Opus 5.5 Specs and Pricing

Model your own workload in the Claude API cost estimator, which includes Opus 5.5, its fast mode, and its cache-read rate.

Claude Opus 5.5 Benchmarks

These are Anthropic’s published launch figures. Opus 5.5 was run with adaptive thinking at max effort unless noted; its Terminal-Bench 4.0 score uses xhigh effort.

Three things are worth reading into this table. First, Opus 5.5 leads Fable 5.1 on every row, at 40% of Fable’s list price. Second, it isn’t a clean sweep: OpenAI’s GPT-6 Astra is ahead on AutomationBench and on agentic scientific research. Third, Anthropic itself cautions that at this level “benchmark margins have become a less reliable guide to real-world differences,” and notes that some scores are pulled down by safeguards routing cybersecurity and biology tasks to other models. Several benchmarks here are newer versions than the ones in Fable 5.1’s September 1 table, so Fable’s numbers differ slightly from our Fable 5.1 guide.

The full head-to-head is in Claude Opus 5.5 vs Fable 5.1, and the cheaper alternative is covered in Sonnet 5.5 vs Opus 5.5.

API Changes: Migrating From Opus 5

Four changes can break code that runs on Opus 5 today. The first three also apply to Fable 5.1:

• Thinking can’t be disabled. Adaptive thinking is always on; control its depth with the effort parameter instead.
• Forced tool use returns an error. tool_choice of type any or tool is rejected — use auto with strict tool use or structured outputs.
• Thinking blocks are tied to the model and the conversation, so editing earlier turns or switching models mid-conversation can invalidate them.
• The older computer_20251124 computer-use tool isn’t accepted on the Claude API and Google Cloud.

One more change alters response shape without failing requests: text Claude writes between tool calls now arrives inside thinking blocks, which are empty at the default display setting. If your app streams that text to users as progress updates, set a display value that returns it, or your UI will go quiet between tool calls.

Availability, Plans, and Claude Code

• Everywhere at launch: the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, plus the Claude apps.
• Bigger subscription limits: alongside the launch, Anthropic raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, and gave subscribers a one-time rate-limit reset they can save and use when they choose.
• Claude Code default: Opus 5.5 is the default model in Claude Code on Pro, Max, Team, Enterprise, and the Anthropic API (Claude Code v2.1.280 or later — run claude update). Fast mode is available there too.
• Zero data retention: available, unlike Fable 5.1, which carries mandatory 30-day retention.

See Is Claude AI Free? for what each plan includes, and Claude Code pricing for how usage limits work in practice.

Safety and Safeguards

• Alignment: Anthropic calls it the strongest-performing model it has tested on its automated behavioral audit, and reports it tried to circumvent containment boundaries about 85% less often than Opus 5 or Claude Mythos 5.1. External testing came from METR and other evaluators before release.
• Cybersecurity: safeguards are on by default and most cybersecurity tasks are re-routed to Claude Opus 4.8. Anthropic says an expansion of its Cyber Verification Program is coming in the following weeks.
• Biology: the same biology safeguards as Fable 5.1, with vetted labs, startups, and pharmaceutical companies able to apply for broader access through the new Life Sciences Verification Program.
• Anti-distillation: preserved-thinking safeguards block editing earlier context on API accounts created on or after August 31, 2026.
• Evaluation awareness: Anthropic notes signs that Opus 5.5 often suspects it is being evaluated, and says building evaluations that catch every failure before deployment remains unsolved.

Where Opus 5.5 Fits in the Lineup

For most people the choice is now between Opus 5.5 and Sonnet 5.5, which costs half as much and lands within two points of Opus 5.5 on several of the same benchmarks. Claude Models Explained covers the full lineup, and the model selector gives a recommendation for your task.

Limitations and Caveats

• Benchmark scores were run at max or xhigh effort, but the API default is medium. Expect lower scores and lower costs at the default; raise effort for hard steps.
• Most cybersecurity work is routed to Opus 4.8 by default, which can surprise security teams until the expanded Cyber Verification Program arrives.
• Four breaking API changes mean migrating from Opus 5 is more than a model-ID swap.
• Fast mode is a research preview and costs double the standard rate.

Frequently Asked Questions