Anthropic released Claude Opus 5.5 on 22 September 2026 with lower API prices and coding benchmark results that the company says exceed those of its more expensive Fable 5.1 model, TechCrunch reported. It is the first model in Anthropic’s 5.5 generation, arriving after Opus 5 was released on 24 July.
Key points
- Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5’s rates.
- Anthropic reports scores of 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode v1.1 Main, ahead of Fable 5.1 on both tests.
- Anthropic says typical workloads cost about 40% less overall because Opus 5.5 uses fewer tokens to complete tasks.
- Anthropic has applied the same class of safeguards used for Fable 5.1 to the new model’s biology and cybersecurity capabilities.
Opus 5.5 changes the token bill
Anthropic set the base API price at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, The New Stack reported. Input tokens are the text sent to a model; output tokens are what it produces. The distinction matters when a task involves many exchanges, because both sides of each exchange contribute to the bill.
Repeated text has its own price. Cache reads fell from $0.50 to $0.20 per million tokens, while cache writes fell from $6.25 to $5. A cache lets a system reuse text it has already supplied, rather like keeping an open reference book on a desk instead of fetching it again for each question. Its price therefore matters most when work repeatedly draws on the same material.
Revising a piece of code could involve repeated questions about the same surrounding text. Reusing that text would cost less at the new cache-read rate, and the total could fall further if the model needed fewer tokens to finish the task. How much it would cost would still depend on the text sent, reused and generated.
Anthropic puts the saving for typical workloads at about 40% overall, according to VentureBeat, because it says Opus 5.5 completes tasks using fewer tokens. The company also says output is generated more than 30% faster. Those are workload claims, distinct from the 20% cut in the published base input and output rates: paying less for each token and using fewer tokens affect a bill in different ways.
Fable 5.1’s base rates are $10 per million input tokens and $50 per million output tokens. Opus 5.5’s corresponding rates are 60% lower, though a task’s cost also depends on how many tokens each model uses. For developers choosing a model to run repeatedly, the price of a request and the work completed by that request are separate questions.
Terminal-Bench 4.0 puts coding claims to work
Anthropic reports a score of 66.4% for Opus 5.5 on Terminal-Bench 4.0, against 55.8% for Fable 5.1 and 52.3% for Opus 5. On FrontierCode v1.1 Main, its published figures are 54.4%, 50.3% and 48.0%, respectively. These are Anthropic’s results on agentic coding evaluations: tasks in which a model works through a software problem across multiple actions rather than merely answering a single coding question.
That distinction is important for software work. An agent may have to inspect code, try a change and respond to what happens before it can finish. A score on a defined set of such tasks records performance under those test conditions; Anthropic itself cautions, according to The New Stack, that a difference of a few benchmark points at this level need not be noticeable in ordinary use.
GitHub tested Opus 5.5 in Copilot CLI and VS Code. Its chief product officer, Mario Rodriguez, said that in VS Code the new model solved more terminal tasks than Opus 5 while taking fewer than half as many steps, TechRepublic reported. That is a comparison from GitHub’s testing in those products, rather than a measure of every coding task a customer might give the model.
Anthropic also tested a larger, sustained job: translating HAProxy from C into Rust. The company says Opus 5.5 and Fable 5.1 both passed nearly all the project’s regression tests, while Opus 5.5 took 9.5 hours rather than 12 and cost 51% less. The test ties its cost claim to a completed piece of work, with the narrower scope of an internal experiment.
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September as lower-priced additions to its own model range, VentureBeat reported. Sol’s base rates are $2 per million input tokens and $10 per million output tokens. Those prices make the amount of work a model completes per request especially relevant to companies comparing systems, rather than leaving the posted token rate as the whole calculation.
Anthropic applies Fable 5.1 safeguards
Anthropic says Opus 5.5 has biology and cybersecurity capabilities comparable to those of its more capable models, and has applied the same safeguards used for Fable 5.1. The controls limit use of the models for tasks including finding exploits in compiled programs and developing recognisable biological weapons. For organisations building on the API, those controls are part of the model’s behaviour alongside its speed and price.
In cybersecurity, general prompts can be routed to Opus 4.8 instead, while access to higher-tier tasks requires entry through Anthropic’s Cyber Verification Program or Life Sciences Verification Program, TechRepublic reported. A system built around Opus 5.5 could therefore receive an answer from another model for a request covered by those controls.
Frontier Design and METR evaluated Opus 5.5 before release. Anthropic said the model frequently recognised that it was being evaluated, complicating its assessment of whether behaviour observed in those tests would carry over to deployment.
