Overview
Claude Opus 5.5 is Anthropic's September 22, 2026 successor to Claude Opus 5. It is the first model in the Claude 5.5 family. Anthropic positions it as a stronger everyday choice for coding, knowledge work, and long-running agents: its published evaluations rise over Opus 5, its API token rates fall, and it generates output faster. The practical upgrade decision is whether those gains outweigh the behavior and safeguard changes in your existing workflow.
Opus 5.5 is available as claude-opus-5-5 on the Claude Platform and across Anthropic's supported platforms, including AWS, Google Cloud, and Microsoft Azure. It is distinct from Claude Fable 5.1, which remains an option for workloads that need its particular capabilities or access policies.
What's New
Stronger Coding and Knowledge Work
In Anthropic's published comparisons, Opus 5.5 scores 66.4% on Terminal-Bench 4.0 versus 52.3% for Opus 5, and 57.8% on CursorBench 4.0 versus 46.6%. GDPval-AA v2.1, a knowledge-work evaluation, rises from 1708 to 1846 Elo. Anthropic also reports gains in computer use and scientific terminal tasks. These are release evaluations, so teams should rerun their own coding, review, and document workflows before forecasting production gains.
Lower Cost per Task and Faster Output
Base API prices fall from $5 to $4 per million input tokens and from $25 to $20 per million output tokens. Cache reads fall from $0.50 to $0.20 per million tokens. Anthropic says Opus 5.5 uses fewer tokens per task and generates output more than 30% faster than Opus 5; at default settings, it estimates 40% lower cost for typical workloads. That 40% is a workload estimate, not a flat rate reduction. A short uncached request, a cache-heavy agent session, and a run that produces substantial thinking tokens will have different bills.
Clearer Reports From Long Tasks
Anthropic says Opus 5.5 puts important information earlier in its answers, follows writing instructions more closely, and reports what it did, found, and needs next after longer work. That makes it worth testing on supervised code review and research tasks where a person must inspect the result. It does not remove the need to check the model's work.
Stronger Safeguards With Different Access
Anthropic reports stronger prompt-injection resistance and its best result to date on its automated behavioral audit. Opus 5.5 also uses safeguards for advanced biology and cybersecurity work similar to those on Fable 5.1. Organizations affected by biology safeguards can apply to the Life Sciences Verification Program now; Opus 5.5 is not yet included in the Cyber Verification Program. This matters when moving a research or security workflow from Opus 5: availability of the model does not guarantee identical handling of every request.
Performance Benchmarks
| Anthropic-reported evaluation | Opus 5.5 | Opus 5 | What to retest |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 52.3% | Multi-step terminal agents |
| CursorBench 4.0 | 57.8% | 46.6% | Ambiguous, multi-file coding |
| GDPval-AA v2.1 | 1846 Elo | 1708 Elo | Professional knowledge work |
| AutomationBench | 40.0% | 26.9% | Connected business workflows |
| OSWorld 2.0, partial | 81.8% | 74.0% | Computer use |
Anthropic reports most Opus 5.5 results at maximum effort, with a different setting for Terminal-Bench 4.0. The figures describe the named evaluation setups, not a guaranteed improvement for a particular application.
Compared With Previous Version
| Decision factor | Opus 5 | Opus 5.5 |
|---|---|---|
| Model ID | claude-opus-5 |
claude-opus-5-5 |
| API input / output per million tokens | $5 / $25 | $4 / $20 |
| Cache reads per million tokens | $0.50 | $0.20 |
| Output speed | Baseline | More than 30% faster, per Anthropic |
| Thinking | Can be disabled under supported effort settings | Cannot be disabled |
| Higher-risk biology and cyber requests | Opus 5 safeguards | Safeguards comparable to Fable 5.1; verify access for affected work |
The headline efficiency gain combines lower unit prices with potentially fewer tokens and turns. Anthropic's separate cost guide recommends measuring the same real tasks on both models because usage patterns determine actual savings.
Migration and Compatibility
- Change the model ID and rerun evaluations. Select
claude-opus-5-5, rerun representative evaluations, and audit the other documented breaking changes: forcedtool_choicevalues (anyor a namedtool) now return 400; on the Claude API and Google Cloud, legacycomputer_20251124integrations must migrate tocomputer_toolset_20260801. - Review thinking settings. Opus 5.5 uses adaptive thinking and does not support thinking switched off. Applications that explicitly disable thinking should update that request path and recheck output-token budgets, latency, and cost.
- Check conversation handling. Anthropic's preserved-thinking model compatibility applies to Opus 5.5 generally; the conversation-prefix integrity check is enforced by default for accounts created on or after August 31, 2026, while older accounts opt in to that check. Integrations that edit previous assistant thinking or conversation prefixes should test Anthropic's preserved-thinking guidance before switching.
- Check restricted workflows. Biology and cybersecurity safeguards can reroute or block affected requests; Anthropic says most cybersecurity tasks are rerouted to Opus 4.8, while routine secure-coding work remains supported. Teams should distinguish the programs: the Life Sciences Verification Program supports eligible Opus 5.5 biology research now, while Anthropic says Opus 5.5 is not yet available through the Cyber Verification Program and that expansion is coming.
- Recalculate fast-mode economics. Fast mode is available as an access-controlled research preview in Claude Code and on the first-party Claude API only, at $8 input and $40 output per million tokens, with up to 2.5× speed. Compare its premium with the standard $4/$20 rate before making it the default.
Who Should Upgrade / Who Should Wait
Upgrade candidates: teams using Opus 5 for code migrations, agentic coding, reviews, knowledge work, or long Claude Code sessions that can benefit from lower cache-read prices. The move is especially attractive when a representative test shows fewer turns or less rework, not just a lower token rate.
Test before switching: API applications that disable thinking, edit past conversation content, depend on predictable output-token budgets, or run biology and cybersecurity tasks subject to verification and fallback policies. Keep Opus 5 as a comparison baseline until those behaviors are measured.
