Anthropic shipped Claude Opus 5.5 on September 22, 2026, the first model in its 5.5 family, and priced it 40% below Opus 5: $4 per million input tokens, $20 per million output, with a fast mode at $8 and $40. It also generates output 30% faster than its predecessor. What's missing is the number every previous flagship launch led with. There's no SWE-bench Verified score anywhere in Anthropic's published benchmark table.
- Opus 5.5 (model ID
claude-opus-5-5) costs $4 input / $20 output per million tokens, a 40% cut from Opus 5, with cache reads at $0.20 and cache writes at $5; fast mode runs $8 input / $40 output. - Anthropic's published scores lean agentic: Terminal-Bench 4.0 at 66.4%, OSWorld 2.0 (computer use) at 81.8%, GDPval-AA v2.1 at 1846 Elo, and no SWE-bench Verified figure at all.
- Xiaomi open-sourced MiMo-V2.6-Pro the same day, a 1-trillion-parameter (42B active) MoE model under MIT license priced at $0.43 input / $0.87 output per million tokens, roughly a tenth of Opus 5.5's input rate.
- Opus 5.5 is live on the Claude Platform, AWS, Google Cloud and Microsoft Azure, with a zero-data-retention option for regulated customers.
What actually changed in Opus 5.5?
Strip away the benchmark question and the release itself is straightforward. Model ID claude-opus-5-5. Pricing drops 40% against Opus 5 across the board: $4 per million input tokens, $20 per million output, $0.20 for cache reads, $5 for cache writes. Fast mode runs $8 and $40. Anthropic also says output generates 30% quicker than Opus 5 did, which matters for agentic workloads where a model calls tools in a loop and every round trip adds latency.
RelatedAnthropic Ships Claude Opus 5 at Half of Fable 5’s Price
Anthropic also claims better resistance to prompt injection, improved scores on its own automated alignment audits, and clearer communication than Opus 5, none of it backed by a published number. The model is live now on the Claude Platform, AWS, Google Cloud and Microsoft Azure, with the zero-data-retention option enterprise and government buyers tend to require.
Why did the benchmark list shift?
This is the part worth sitting with. Opus 5.5's card runs through Terminal-Bench 4.0 (66.4%), FrontierCode v1.1 (54.4%), CursorBench 4.0 (57.8%), AutomationBench (40.0%), OSWorld 2.0 for computer use (81.8%) and Chartography for visual reasoning (89.0%), plus GDPval-AA v2.1 at 1846 Elo. What isn't there is SWE-bench Verified, the "can it patch a real GitHub issue" test that anchored nearly every frontier coding-model launch since 2024.
Anthropic didn't say anything about SWE-bench, and nothing suggests it was run and withheld. It just isn't in the headline table this time. But look at what replaced it. Terminal-Bench checks whether a model can operate a shell across a long session. OSWorld checks whether it can drive a real desktop through real applications. AutomationBench and CursorBench track how developers actually work inside an editor or a workflow tool, not against one isolated diff. GDPval grades output the way a professional reviewer would, task by task, reported as a rating rather than a pass rate. Together, the set measures whether a model can carry a job to completion across many steps, not whether it can produce one correct patch. That's the question buyers evaluating an agent platform should actually be asking.
How exposed is the open-weights competition?
Anthropic wasn't the only lab with news on September 22. Xiaomi open-sourced MiMo-V2.6-Pro the same day: a mixture-of-experts model with 1 trillion total parameters, 42 billion active per token, released under an MIT license. Artificial Analysis scored it 46 on its Intelligence Index, the top mark of any open-weights model tracked so far, at $0.43 input / $0.87 output per million tokens, roughly a ninth of Opus 5.5's input rate.
That gap doesn't erase Opus 5.5's advantages: enterprises buying a hosted, audited, zero-retention model aren't shopping the same aisle as a team standing up open weights on its own infrastructure. But a frontier-class open model landing at a fraction of the price, the same day, is the kind of coincidence that comes up in procurement conversations anyway. We covered MiMo-V2.6-Pro separately in a companion piece.
RelatedQwen3.8-27B ships open weights, scoreboard attached
What it means for the market
Anthropic is private, so there's no ticker to watch directly. Exposure runs through the companies reselling Opus 5.5: Amazon (AMZN) via AWS, Alphabet (GOOGL) via Google Cloud, and Microsoft (MSFT) via Azure all list it as a hosted option. A 40% price cut helps adoption, but it's also a data point on where per-token margins across the AI stack are headed, especially with a $0.43-input open model landing the same week. The signal for investors isn't about Anthropic's own valuation. It's about how much pricing power the infrastructure layer keeps as inference gets cheaper and open weights close the quality gap. Worth watching: how the hyperscalers describe AI-driven cloud consumption in coming earnings calls, and whether rival labs answer with price cuts of their own.
What about the safety story?
Opus 5.5 ships with safeguards comparable to Claude Fable 5.1's, covering cybersecurity, biology and model distillation, plus a preserved-thinking protection aimed at making the model harder to distill into a smaller copycat. Two vetted-access programs launch alongside it: a Life Sciences Verification Program for biology researchers and a Cyber Verification Program for security practitioners, gating higher-risk capability behind identity checks rather than an open API key.
| Metric | Opus 5.5 | Opus 5 | MiMo-V2.6-Pro |
|---|---|---|---|
| Input $/M tokens | $4.00 | Not restated; Opus 5.5 is a 40% cut | $0.43 |
| Output $/M tokens | $20.00 | Not restated; Opus 5.5 is a 40% cut | $0.87 |
| Fast mode in/out | $8 / $40 | Not offered | Not applicable |
| Cache read / write | $0.20 / $5 | Not stated here | Not stated |
| Architecture | Not disclosed | Not disclosed | MoE, 1T total / 42B active |
| License / access | Hosted: Claude Platform, AWS, GCP, Azure | Same channels | Open weights, MIT license |
| Headline evals published | Terminal-Bench, OSWorld, GDPval, no SWE-bench | SWE-bench era suite | Artificial Analysis Intelligence Index: 46 |
- Whether SWE-bench comes back. If Anthropic or rivals reintroduce a SWE-bench-style score in the next model card, that tells you this launch's omission was tactical rather than a lasting shift in what gets measured.
- Hyperscaler earnings language. AWS, Google Cloud and Azure results over the next two quarters should show whether Opus 5.5's price cut moved usage volume or just compressed margin on existing traffic.
- Whether other labs match the cut. A 40% reduction from the market's highest-profile lab puts pressure on every competitor's next pricing update.
- Uptake on the verification programs. How many biology and cybersecurity researchers actually enroll says something about whether gated access is a real safety layer or a formality.
Our take
The pricing and speed numbers are the easy story here, and they're genuinely good for anyone building on the Claude Platform. But the benchmark table is the more interesting document. Two years of frontier launches trained everyone, us included, to check the SWE-bench line first. Opus 5.5's card doesn't have one, and the replacement set reads like Anthropic is grading for a different job: not "can this model write a correct patch" but "can it sit down at a computer and finish the task." That's closer to what agentic products actually need, and harder to game than a static benchmark trained against for two years. Whether the rest of the industry follows that shift, or keeps reaching for the old scorecard out of habit, is the open question.
- Official Anthropic: Claude Opus 5.5 announcement, pricing table and published benchmark scores
- Benchmark Artificial Analysis: Claude Opus 5.5 independent third-party model analysis
- Reference GenZTech AI Coding Leaderboard our live tracker of coding-model benchmark results
Original analysis by GenZTech, based on Anthropic's published model card and pricing.