Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens. It says the model matches its top Fable 5.1 on most work, and one independent index ranks it first.
New to this? Read it in simple words
- Anthropic released Claude Opus 5.5 on September 22. It costs $4 per million input tokens, 20 percent less than Opus 5.
- Anthropic says it performs like its more expensive Fable 5.1 on most work. In its highest-effort setting, one independent ranking puts it first of 206 models.
- Lower prices matter only if the quality stays good on your own tasks. When a safeguard stops a task, an older model can finish it.
- Most other figures come from Anthropic. None of the pages we read showed the outside testers’ own findings.
- Token
- A small piece of text, often part of a word. AI use is priced per token.
- Safeguard
- A rule or tool that protects against harm or misuse.
Lower prices and a faster mode
Claude Opus 5.5 is available through Anthropic’s apps and API, and on Amazon Web Services, Google Cloud, and Microsoft Azure. Developers call it with the name claude-opus-5-5.
The price is $4 per million input tokens and $20 per million output tokens. Opus 5 cost $5 and $25. A token is a small piece of text, often part of a word. Reading stored text from the cache now costs $0.20 per million tokens instead of $0.50.
A fast mode runs up to 2.5 times faster and costs $8 and $40 per million tokens. Anthropic says the model is about 40 percent cheaper than Opus 5 on typical work, but that figure depends on its own mix of tasks. Its Claude blog says usage limits on paid plans now go about 25 percent further.
Opus 5.5 costs $4 and $20 per million tokens. Most performance figures come from Anthropic.
Strong results on Anthropic’s tables, and one outside ranking
Anthropic reports a score of 66.4 percent on Terminal-Bench 4.0, a test of tasks that an AI agent completes in a computer terminal. It lists Fable 5.1 at 55.8 percent, Opus 5 at 52.3 percent, and OpenAI’s GPT-6 Astra at 57.9 percent, as reported by OpenAI.
Anthropic’s own table also shows GPT-6 Astra ahead on two tests, AutomationBench and Terminal-Bench-Science. Anthropic itself says that small benchmark gaps are now a weak guide to real quality.
Artificial Analysis, an independent company that compares models, gives Opus 5.5 in its highest-effort setting a score of 58 on its Intelligence Index. That ranks it first of the 206 models it tracks.
Safety claims, and tasks handed to older models
Outside groups, including METR and Frontier Design, tested the model before release. Anthropic says it scored better than any earlier model on its automated behaviour audit of about 2,000 scenarios. It also admits that the model often suspects it is being tested.
Anthropic rates the model’s biology and cyber abilities as similar to Mythos 5.1, so it uses stronger safeguards. Most cybersecurity requests go to the older Opus 4.8. When a safeguard stops a task, Opus 4.8 or Opus 5 can complete it instead.
Yahoo Finance notes that this is Anthropic’s first new model since chief executive Dario Amodei called for slowing frontier AI. None of the pages we read included METR’s or Frontier Design’s own findings. Teams should test Opus 5.5 on their own work and check when a task moves to an older model.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3Research · September 22, 2026Anthropic releases Opus 5.5 with lower prices and Fable-level performance TechCrunch
- 4Research · September 22, 2026Claude Opus 5.5 (max with fallback): Intelligence, Performance & Price Analysis Artificial Analysis
- 5Research · September 22, 2026Anthropic launches Opus 5.5, its first model since CEO Amodei called for AI slowdown Yahoo Finance
We read Anthropic’s launch post and pricing blog, then compared them with TechCrunch, Yahoo Finance, and the independent Artificial Analysis ranking. Apart from that ranking, the performance, cost, and safety figures come from Anthropic, so we label them as company claims.