The Main Change Is Economic
Anthropic released Claude Opus 5.5 on 22 September 2026, with the model id claude-opus-5-5. The release improves several reported benchmark results over the company's previous flagship while also reducing API prices. That combination is more important than another isolated benchmark lead because it changes the cost of deploying a frontier model, not just the scores attached to it.
VentureBeat reported standard API pricing of $4 per million input tokens and $20 per million output tokens, down 20 percent from Opus 5. Anthropic also says Opus 5.5 generally needs fewer tokens to complete a task. Quartz reported that Anthropic puts the resulting cost reduction for typical workloads at roughly 40 percent overall compared with Opus 5.
For developers and engineering managers, this makes Opus 5.5 a cost-performance release rather than simply a capability update. A stronger model is easier to justify when its unit price moves down instead of up. Buyers still need workload-specific testing, but the economic question has changed: higher-end performance no longer automatically arrives with higher per-token pricing. It is a different shape of release from the one the AI model race produced through most of 2026.
Pricing Goes Beyond Input and Output Tokens
The headline rates are only part of the pricing change. VentureBeat reported that cache writes fall from $6.25 to $5 per million tokens, while cache reads drop from $0.50 to $0.20 per million tokens. The cache-read reduction is 60 percent, making repeated use of cached material notably cheaper under the new schedule.
Anthropic is also offering a faster variant at $8 per million input tokens and $40 per million output tokens. That gives teams a separate price point when speed is important, rather than making the faster option the default cost for every deployment. Anthropic has not published a single universal cost figure that applies across every workload.
The practical issue for buyers is therefore broader than comparing one input rate and one output rate. Token usage, output length, cache behaviour and the choice of standard or faster service can all affect the bill. Quartz's report of roughly 40 percent lower typical workload cost is Anthropic's estimate, not a fixed discount that every application will reproduce.
How Opus 5.5 Compares on Benchmarks
The reported results show Opus 5.5 ahead of Anthropic's previous flagship, Claude Fable 5.1, on several tests. VentureBeat reported 66.4 percent on Terminal-Bench 4.0 against 55.8 percent for Fable 5.1, 54.4 percent against 50.3 on FrontierCode v1.1 Main, 57.8 percent against 51.8 on CursorBench 4.0, and 1846 Elo against 1735 on GDPval-AA v2.1.
The comparison with OpenAI's GPT-6 Astra is also mixed rather than one-sided. OfficeChai reported Opus 5.5 ahead on Terminal-Bench 4.0, FrontierCode v1.1, GDPval-AA v2.1, Humanity's Last Exam with tools, and OSWorld 2.0. Astra leads on AutomationBench and Terminal-Bench-Science. All benchmark numbers discussed here are vendor-published unless stated otherwise, a caveat we apply to every model launch in our AI benchmarks coverage.
| Benchmark | Claude Opus 5.5 | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | Fable 5.1: 55.8%; GPT-6 Astra: 57.9% |
| FrontierCode v1.1 Main | 54.4% | Fable 5.1: 50.3%; GPT-6 Astra: 53.3% |
| CursorBench 4.0 | 57.8% | Fable 5.1: 51.8% |
| GDPval-AA v2.1 | 1846 Elo | Fable 5.1: 1735; GPT-6 Astra: 1542 |
| Humanity's Last Exam with tools | 67.7% | GPT-6 Astra: 57.2% |
| Terminal-Bench-Science | 58.7% | GPT-6 Astra: 64.6% |
| AutomationBench | Score not published | GPT-6 Astra is ahead |
| OSWorld 2.0 | Ahead of GPT-6 Astra | Exact scores not published |
Those results matter most when read alongside the pricing. VentureBeat also reported that Fable 5.1 remains roughly 150 percent more expensive per API token. Opus 5.5 therefore improves the listed benchmark results over that previous flagship while occupying a substantially lower API price position.
Where Opus 5.5 Does Not Lead
The release does not establish Opus 5.5 as the leader on every workload. OfficeChai reported that GPT-6 Astra remains ahead on AutomationBench, a Zapier-managed test of business workflows. Exact AutomationBench scores have not been published, so the size of that gap cannot be quantified.
Astra also leads on Terminal-Bench-Science, scoring 64.6 percent against 58.7 percent for Opus 5.5. That is a useful counterweight to the results where Anthropic's model leads. The two models do not have the same relative position across every coding, workflow and science-oriented evaluation.
For engineering teams, that argues against selecting a model from an aggregate impression of benchmark performance. A workload centred on business automation or scientific terminal tasks faces a different published comparison from one centred on the tests where Opus 5.5 leads. The sensible evaluation target is the task a team actually intends to run.
Context, Output Limits and Deployment Reach
Opus 5.5 has a 1 million token context window. VentureBeat reported a maximum synchronous output of 128,000 tokens and a maximum batch output of 300,000 tokens. Those limits make both input capacity and output capacity material parts of the model's specification, rather than treating long context as an input-only feature.
One early user reportedly completed a 680,000-line code migration in less than a day. That is a single reported customer result, not evidence that other migrations of similar size will finish on the same schedule. Its value is as an example of the kind of large engineering task being attempted with the model, not as a general performance guarantee.
Availability is broad from launch. VentureBeat reported access through the Anthropic API, AWS, Google Cloud and the Microsoft ecosystem. For organisations that already standardise on one of those routes, model evaluation does not necessarily require adopting a new infrastructure provider. Exact deployment choices will still depend on each organisation's existing stack and requirements. (For the consumer product built on these models, see our Claude review and Claude Code review.)
Anthropic also describes stronger cybersecurity safeguards and a reduced risk of containment bypass. External testing took place before release through groups including METR. Those claims are relevant to teams evaluating a frontier model for sensitive development work, although they do not remove the need for an organisation's own controls and deployment review. Our running Claude coverage tracks how that line-up has changed through 2026.
What to Check Before Switching
Lower pricing and stronger benchmark results make Opus 5.5 worth testing, but neither tells an engineering team whether an existing workload should move immediately. The most useful comparison is against the current system using representative prompts, expected output sizes, cache patterns and the specific tasks that matter in production.
- Measure actual input and output token use on representative jobs.
- Check how much of the workload can benefit from $0.20 per million token cache reads.
- Compare the standard $4/$20 pricing with the $8/$40 faster variant only where additional speed matters.
- Test the tasks that resemble your own workload instead of relying on one benchmark.
- Include AutomationBench and Terminal-Bench-Science results in the decision where those task types are relevant.
- Verify whether the 1 million token context and output limits change how jobs can be structured.
- Confirm that the required deployment route is available through Anthropic API, AWS, Google Cloud or the Microsoft ecosystem.
The economic case should also be measured at the workload level. Anthropic's roughly 40 percent typical cost reduction versus Opus 5 includes both lower prices and lower token use. Teams should therefore compare complete task cost, not assume that the 20 percent reduction in headline token pricing describes the full savings or guarantees the same result for every application.
A Week of Pressure on Price-Performance
Opus 5.5 arrives in a week when several frontier model releases are pushing price-performance from different directions. xAI released Grok 4.7 on 21 September 2026 while keeping API pricing at $2 per million input tokens and $6 per million output tokens. The release also posted gains on CursorBench and DeepSWE.
Around 21 to 22 September, Xiaomi open-sourced MiMo-V2.6-Pro under an MIT licence. It is a roughly 1.02-trillion-parameter mixture-of-experts model that activates about 42 billion parameters per token, supports a 1 million token context window, and scores about 46 on the Artificial Analysis Intelligence Index. That puts another large open model into the same week's discussion about capability and deployment economics, extending a pattern we covered in the China versus US model race.
The significance of Anthropic's release is therefore not simply that another flagship posts higher scores. Opus 5.5 combines lower per-token pricing, an asserted reduction in the number of tokens needed per task, and stronger reported results than Anthropic's previous flagship on the listed tests. At the same time, Grok 4.7 applies price pressure from a lower-priced closed API, while Xiaomi applies a different form of pressure through an MIT-licensed model.
For developers, managers and buyers, this is the more consequential shift. Model evaluation is moving further away from a simple choice between paying more for the frontier or accepting lower capability to save money. Opus 5.5 does not lead every benchmark, but it shows that a new flagship can improve published performance while its standard API price falls. That makes cost per completed workload, rather than model generation alone, an increasingly important comparison.
FAQ
How much does Claude Opus 5.5 cost?
Standard API pricing is $4 per million input tokens and $20 per million output tokens, 20 percent below Opus 5. Cache writes cost $5 per million tokens and cache reads $0.20 per million. A faster variant costs $8 per million input tokens and $40 per million output tokens.
How large is the Claude Opus 5.5 context window?
Claude Opus 5.5 has a 1 million token context window. VentureBeat reported a maximum synchronous output of 128,000 tokens and a maximum batch output of 300,000 tokens. Those limits matter for teams planning large coding, document or batch workloads because both input capacity and maximum generated output need to fit the job.
Is Claude Opus 5.5 better than Fable 5.1?
On the reported benchmarks, Opus 5.5 scores higher than Fable 5.1 on Terminal-Bench 4.0, FrontierCode v1.1 Main, CursorBench 4.0 and GDPval-AA v2.1. VentureBeat also reported that Fable 5.1 remains roughly 150 percent more expensive per API token. Exact real-world performance will still depend on the workload being tested.
Does Claude Opus 5.5 beat GPT-6 Astra?
The published comparison is mixed. Opus 5.5 leads GPT-6 Astra on Terminal-Bench 4.0, FrontierCode v1.1, GDPval-AA v2.1, Humanity's Last Exam with tools and OSWorld 2.0. Astra leads on AutomationBench and Terminal-Bench-Science, where it scores 64.6 percent against 58.7 percent for Opus 5.5.
Where is Claude Opus 5.5 available?
Claude Opus 5.5 is available through the Anthropic API, AWS, Google Cloud and the Microsoft ecosystem. That gives organisations several deployment routes without tying access to Anthropic's own API alone. The model id for API use is claude-opus-5-5.
What else changed in the AI model market this week?
xAI released Grok 4.7 on 21 September at $2 per million input tokens and $6 per million output tokens, while Xiaomi open-sourced MiMo-V2.6-Pro under an MIT licence around 21 to 22 September. Together with Opus 5.5 pricing, those releases put additional attention on price-performance rather than benchmark improvement alone.
