Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in the Claude 5.5 family. The company says the new model reaches Claude Fable 5.1-level performance on most work while costing 40% less to run than Opus 5 on typical workloads, and generates output more than 30% faster.
For people who use AI to code, research, analyze data, or produce business documents, that combination matters more than a small benchmark gain. Below: what changed, what the launch numbers mean and where they are unverified, what Opus 5.5 costs, and who should use it. We have not tested the model ourselves.
Claude Opus 5.5 at a glance
- Release date: September 22, 2026
- Best suited for: complex coding, long-running agent tasks, research, financial analysis, and professional knowledge work
- API model name: claude-opus-5-5
- Standard API price: $4 per million input tokens and $20 per million output tokens
- Speed: Anthropic reports output generation more than 30% faster than Opus 5
- Availability: Claude apps, Claude Code, the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure
- Coming next: Sonnet 5.5 and Haiku 5.5 in the coming weeks, no dates or prices
What is new in Claude Opus 5.5?
1. Stronger agentic coding
The biggest improvements are aimed at work that takes many steps: understanding a large codebase, planning a migration, editing multiple files, running tests, checking results, and correcting mistakes without losing the original goal.
Anthropic says an early tester completed a 680,000-line code migration in less than a day, and another audited and fixed a 200,000-line codebase in under three hours where Opus 5 took more than 20 hours and 2.5 times the tokens. These are selected early-access examples, not controlled studies. Testers also reported fewer repeated attempts, fewer tool calls, and more complete edits, which makes an agent easier to supervise because there are fewer partial changes to review.
2. Better performance per dollar, with a caveat
Per token, Opus 5.5 is 20% cheaper than Opus 5, and cache reads, which Anthropic says make up most of the cost of coding agents, are 60% cheaper.
| Price per million tokens | Claude Opus 5.5 | Claude Opus 5 | Claude Fable 5.1 |
|---|---|---|---|
| Input | $4.00 | $5.00 | $10.00 |
| Output | $20.00 | $25.00 | $50.00 |
| Cache writes (5 minutes) | $5.00 | $6.25 | $12.50 |
| Cache reads | $0.20 | $0.50 | $0.25 |
| Fast mode, input and output | $8 and $40, up to 2.5x speed | Available | Not listed |
Source: Claude Platform pricing docs, September 22, 2026. Against Fable 5.1, input, output, and cache-write rates are 60% lower; cache reads are 20% lower because Fable 5.1 already had a discounted read rate. Our Claude pricing guide lists every model and will be updated for the 5.5 family.
The 40% figure is a separate claim: Anthropic’s estimate for typical workloads at default effort, combining the price cut with fewer tokens per task. Customers quoted in the announcement report token reductions between 20% and 66%, as percentages rather than reproducible counts.
The one independent data point does not confirm it yet. Artificial Analysis ran Opus 5.5 at max effort and recorded 260 million output tokens across its Intelligence Index suite, against a median of 92 million and about 190 million for Fable 5.1. Anthropic’s claim concerns default effort, so the two do not contradict each other directly. But lower unit prices make Opus 5.5 cheaper only at equal token usage, and at the highest setting it is the most verbose frontier model the firm has measured.
3. Clearer, less verbose communication
Anthropic says Opus 5.5 puts the most important information first, uses less jargon, and follows the writing rules you give it, addressing the most common complaint about Opus 5. Shorter explanations are easier to verify, and a clear summary makes it faster to approve or correct the model’s next step. For everyday users this may be the most noticeable upgrade; there is no independent test of it yet.
4. Stronger safety results, and Anthropic’s own caveats
Anthropic reports its best results so far on an automated behavioral audit of nearly 2,000 scenarios. In a new evaluation of attempts to cross containment boundaries, Opus 5.5 tried about 85% less often than Opus 5 or Claude Mythos 5.1, with every attempt low severity and self-reported, and it is more resistant to prompt injection than Opus 5. METR and Frontier Design evaluated it before release, the first launch since Anthropic’s chief executive, Dario Amodei, called on September 12, 2026, for pacing AI progress so safety practices stay ahead.
Three caveats from the same announcement matter in practice:
- Anthropic says Opus 5.5 often suspects it is being evaluated, which limits how well pre-release tests predict real-world behavior, and that building evaluations that catch every failure before deployment remains unsolved.
- Because its biology and cybersecurity capabilities are comparable to Mythos 5.1, the model ships with Fable 5.1-class safeguards: most cybersecurity tasks are re-routed to Opus 4.8, and biology research requires the Life Sciences Verification Program. Our Fable 5.1 and Mythos 5.1 explainer covers that access model.
- Opus 5.5 is no longer available with thinking mode switched off. Pipelines that ran Opus 5 without thinking to save tokens lose that option.
Claude Opus 5.5 benchmark results
Anthropic’s launch table shows gains across coding, computer use, and knowledge work.
| Benchmark | What it measures | Opus 5.5 | Opus 5 | Fable 5.1 |
|---|---|---|---|---|
| Terminal-Bench 4.0 (±2.6 pts) | Agentic terminal coding | 66.4% (xhigh) | 52.3% | 55.8% |
| FrontierCode v1.1 | Merge-ready agentic code changes | 54.4% | 48.0% | 50.3% |
| CursorBench 4.0 | Ambiguous, multi-file coding tasks | 57.8% | 46.6% | 51.8% |
| GDPval-AA v2.1 | Real-world work across 44 occupations | 1,846 Elo | 1,708 Elo | 1,735 Elo |
| Terminal-Bench-Science 0.1 (±3.5 to 5 pts) | Agentic scientific research | 58.7% | 29.0% | 52.6% |
| OSWorld 2.0, “partial” (not defined) | Computer-use tasks | 81.8% | 74.0% | 80.7% |
These numbers need context. Unless marked, Opus 5.5 scores use adaptive thinking at max effort, not the default setting behind the cost claim. Competitor figures in Anthropic’s full table (GPT-6 Astra 57.9% on Terminal-Bench, 53.3% on FrontierCode) are OpenAI’s reported numbers. The error bars put several gaps to Fable 5.1 within noise. Where safeguards intervened, fallback models completed the task, so some rows score a production system rather than Opus 5.5 alone. Anthropic itself cautions that benchmark margins are a less reliable guide to real differences at this level and that the gap to Fable 5.1 is narrower than the table suggests.
The more useful signal may be the cost curve. On FrontierCode, Anthropic reports that Opus 5.5 at default medium effort scores 54.6%, above GPT-6 Astra’s published top score of 53.3%, at roughly one-fifth of the cost per task. Treat that as a vendor-reported result. On Artificial Analysis’s independent index, Opus 5.5 scores 58, first of the 206 models in its comparison set, against 53 for Fable 5.1 and 53 for GPT-6 Astra.
What Claude Opus 5.5 means for everyday users
You do not need to be a developer to benefit. Anthropic says 16 of 18 research reports produced by Opus 5.5 passed an internal quality bar that rejected any invented number or quotation; Fable 5.1 and Opus 5 did not pass in that test. In a merger-analysis test, Opus 5.5 built the financial model and executive presentation in 63 minutes against 93 for Opus 5, at half the cost.
For subscribers, Anthropic is raising five-hour usage limits on Pro, Max, and Team plans by an amount it did not state, and giving subscription users a rate-limit reset that can be saved and used later, with mechanics not yet documented. Which plans get Opus 5.5 as the default model was also not stated on September 22.
Good uses for Opus 5.5 include:
- Reviewing a large document set and producing a sourced briefing
- Building a financial model and turning the result into an executive presentation
- Planning and executing a codebase-wide migration
- Investigating a difficult bug across several services or repositories
- Running a multi-step workflow that uses tools, websites, and connected apps
- Rewriting complex material into a clear report that follows a defined style guide
Should you switch to Claude Opus 5.5?
Choose Opus 5.5 when the cost of an incomplete or incorrect result is higher than the cost of the model: complex engineering work, consequential research, long agent sessions, or deliverables that would otherwise need several rounds of revision. A smaller, cheaper model remains the better choice for quick summaries, simple rewriting, classification, or high-volume routine tasks.
- If you use Opus 5 for difficult work through the API: the upgrade looks straightforward, but verify it. Run your real tasks at the default effort setting, where Anthropic’s savings claim applies, and compare total billed tokens and cost per completed task before changing production defaults.
- If you use Fable 5.1: unit rates are 60% lower and Anthropic says most work lands at Fable 5.1’s level. Test your hardest tasks before assuming parity, and check whether losing thinking-off mode affects any pipeline.
- If you use Claude only for short chats: the improved writing style may be more noticeable than the raw capability gain.
The bigger picture
Opus 5.5 is a release about efficiency as much as intelligence. Whether the efficiency claim holds outside Anthropic’s own tests is the open question. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks; if they inherit the same gains, the biggest impact of the 5.5 family may come from the lower-cost everyday models. Our Opus 5 guide remains the reference for the model most plans default to until then.