Claude Opus 5.5: What's New and What Changes
Claude Opus 5.5 launched on September 22, 2026 as the first model in the Claude 5.5 family. Anthropic positions it for long-running agentic coding and knowledge work, and says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Claude Opus 5 on typical workloads.
Quick overview
- Model ID:
claude-opus-5-5 - Amazon Bedrock:
anthropic.claude-opus-5-5 - Google Cloud / Microsoft Foundry:
claude-opus-5-5 - Context window: 1M tokens
- Maximum output: 128K tokens
- Batch maximum output: 300K tokens with the required beta header
- Input / output price: $4 / $20 per million tokens
- 5-minute cache write: $5 / MTok
- 1-hour cache write: $8 / MTok
- Cache read: $0.20 / MTok
- Thinking: adaptive and always on
- Default effort:
medium - Reliable knowledge cutoff: June 2026
- Input / output modalities: text and images → text
What changed in this release
Better long-running coding
Anthropic's release materials focus on codebase-wide migrations, audits, difficult debugging, and agents working across multiple repositories. Early examples include a 680,000-line migration completed in less than a day and an audit-and-fix run across a 200,000-line codebase.
These are Anthropic and partner reports, not substitutes for your own evals. They do show the target: Opus 5.5 is intended to close a complete loop with fewer intermediate steps, fewer repeated tool calls, and less supervision.
Better token efficiency
Compared with Opus 5, Opus 5.5 reduces base input and output prices by 20% and cache-read pricing by 60%. Anthropic estimates that typical workloads cost about 40% less to run, while output is generated more than 30% faster.
Cache reads matter especially for Claude Code. Long sessions repeatedly read project context, system instructions, and tool definitions. Compare cache hit rate, tool-call count, and total tokens per successful task—not just input and output prices.
Clearer communication
Anthropic says Opus 5.5 puts the important information earlier, uses less unnecessary jargon, and follows writing rules more closely. For long sessions, clearer plans, progress updates, and error explanations make the work easier to review.
Stronger safety boundaries
Anthropic reports that Opus 5.5 achieved its strongest results so far on its automated behavioral audit and improved defenses against prompt injection, out-of-scope behavior, and hard-to-reverse actions. Because its biology and cybersecurity capability is comparable to Mythos 5.1, those areas use safeguards similar to Fable 5.1. Vetted researchers can apply for the Life Sciences Verification Program, while cybersecurity access is managed through the Cyber Verification Program.
New platform capabilities
Opus 5.5 supports:
- per-message effort (beta);
- mid-conversation system messages;
- task budgets;
- prompt caching with a 512-token minimum cacheable prompt;
- the Batch API;
- the Files API and PDF support;
- vision;
- server-side and client-side tools;
- on-demand compaction (beta).
Fast mode is also available as a Claude API research preview. It is currently limited to the Claude API, is not available on Bedrock, Claude Platform on AWS, Google Cloud, or Microsoft Foundry, and uses a separate beta header and pricing.
Key signals from the official comparison
Anthropic's launch evaluation reports strong Opus 5.5 results across agentic coding, knowledge work, and computer use:
| Evaluation | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — |
| GDPval-AA v2.1 | 1846 Elo | 1735 Elo | 1708 Elo | 1542 Elo |
| OSWorld 2.1, partial | 81.8% | 80.7% | 74.0% | — |
These are official launch comparisons. Some rows use different effort settings, and production safeguards can route sensitive tasks to another model. Treat the table as directional evidence, then test the exact tasks and harness you plan to deploy.
When should you upgrade?
Start with:
- codebase-wide work that required repeated prompting with Opus 5;
- long-running agents with many tool calls and intermediate checks;
- research that needs fact verification across documents, web pages, and code;
- code, spreadsheets, reports, or presentations that need to be produced as finished artifacts;
- production work where failure is expensive but Fable 5.1 pricing is hard to justify.
Important caveat
Do not treat Opus 5.5 as a string replacement for Opus 5. Its default effort changes from Opus 5's high to medium, while the model may think more at the same effort setting. Re-run an effort sweep and leave enough max_tokens headroom for thinking. Thinking cannot be disabled, and old thinking blocks, tool-choice settings, and computer-use configurations may need migration.
Next, read Migrate from Opus 5 to Opus 5.5.