Claude Opus 5.5
Introducing Claude Opus 5.5
September 22, 2026
Anthropic is proud to unveil Claude Opus 5.5, the premiere model of our latest Claude 5.5 lineage. This model represents a significant leap in efficiency: it delivers performance comparable to Claude Fable 5.1 across most tasks while reducing operational costs by compared to Opus 5.
This release follows our commitment to "pacing the frontier." To ensure safety and reliability, Opus 5.5 underwent rigorous pre-release scrutiny by external partners, including METR and Frontier Design.
"It writes the way I do." — Early Beta Tester
🛡️ Safety and Alignment
Opus 5.5 has achieved the highest marks to date on our automated behavioral audit, our most stringent alignment protocol.
Key Safety Enhancements:
- Boundary Adherence: Significantly lower probability of executing irreversible actions or exceeding defined constraints.
- Robustness: Enhanced resistance to
prompt injectionattacks compared to Opus 5. - Expanded Testing: We have broadened our evaluation scope to include:
- Extended-duration tasks.
- Logically impossible requests.
- Simulations based on historical real-world incidents.
For a deep dive into our methodology, please refer to the Opus 5.5 System Card.
Specialized Access
Due to its high proficiency in cybersecurity and biology (matching Claude Mythos 5.1), we are implementing tiered safeguards:
- Biology: Available via the Life Sciences Verification Program.
- Cybersecurity: Access is expanding via the Cyber Verification Program for verified practitioners.
🚀 Performance Breakthroughs
Opus 5.5 isn't just a marginal update; it is a new industry leader. Early adopters have reported massive productivity gains in complex workflows.
Real-World Impact
- Massive Code Migration: One user successfully migrated
680,000 lines of codein under 24 hours—a project that typically requires several weeks of human engineering. - Software Optimization: In a test to reduce web app load times, Opus 5.5 succeeded in 39 out of 40 instances. In contrast,
Opus 5often made changes that inadvertently altered the application's behavior. - Creative Polish: In game development prompts, Opus 5.5 outperformed all other models in visual fidelity and overall polish.
The "Work Partner" Experience
Users report that the writing style is clearer and more intuitive. By prioritizing the most critical information at the start of a response, it functions as a more effective collaborator during marathon sessions.
📉 Economics and Efficiency
We have optimized the compute requirements for Opus 5.5, allowing us to pass those savings directly to our users.
Cost Comparison
The cost reduction can be expressed as:
| Metric | Opus 5.5 Price | Comparison to Opus 5 |
|---|---|---|
| Input Tokens | $4 / million | 20% Lower |
| Output Tokens | $20 / million | 20% Lower |
| Cache Reads | $0.20 / million | 60% Lower |
Additional Benefits:
- Speed: Generation is faster than its predecessor.
- Capacity: Increased five-hour usage windows for Pro, Max, Team, and Enterprise tiers.
- Flexibility: Subscription users now receive a rate limit reset that can be stored and triggered manually.
📊 Benchmarks & Technical Analysis
While Opus 5.5 leads in knowledge work, computer use, and agentic coding, we note that at this level of intelligence, small benchmark margins are less indicative of real-world experience. The practical gap between Opus 5.5 and Fable 5.1 is narrower than the numbers suggest.
Comparative Performance Table
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (Agentic Coding) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 (Agentic Coding) | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 (Agentic Coding) | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 (Knowledge Work) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (Business Workflows) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam (Reasoning) | 67.7%* | 65.6%* | 63.6%* | 57.2%* | — |
| Terminal-Bench-Science 0.1 (Research) | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 (Computer Use) | 81.8% (p) | 80.7% (p) | 74.0% (p) | — | — |
| Chartography (Visual Recognition) | 89.0%* | 88.4%* | 83.4%* | — | — |
*Results include tool use. (p) denotes partial completion.
Technical Implementation Notes
- Adaptive Thinking: All Opus 5.5 results were achieved using "max effort" adaptive thinking.
- Safeguard Impact: Because certain high-risk tasks were routed to Opus 5 or 4.8, the benchmark scores for those specific categories may be slightly lower than the model's raw potential.
- Error Margin: For Terminal-Bench 4.0, the standard error is pts for Opus 5.5.
📅 What's Next?
The rollout of the 5.5 family continues shortly.
- Claude Sonnet 5.5 (Coming soon)
- Claude Haiku 5.5 (Coming soon)
These upcoming models will inherit the same advancements in safety, speed, and overall performance.