Opus 5.5, released September 22, looks like a flagship but is really an accounting move
Anthropic shipped Claude Opus 5.5 on September 22. The pricing tells you it is not here to show off: 4 dollars per million input tokens and 20 dollars per million output, about 20 percent cheaper per token than Opus 5 and roughly 40 percent cheaper to run overall. Cache reads dropped from 0.50 to 0.20 dollars per million, a 60 percent cut, and long context is exactly where agents burn the most money. Anthropic itself says typical workloads at default settings run about 40 percent cheaper than Opus 5.
Benchmarks: hugging Fable 5.1, not beating it
On Terminal-Bench 4.0, the long-horizon agentic coding test, Opus 5.5 scores 66.4 percent at xhigh effort, above GPT-6 Astra's 57.9 and its own Fable 5.1's 55.8. Two caveats: these are Anthropic's own published numbers, and they are at maximum effort; the default-effort score is lower. On FrontierCode v1.1 it posts 54.4 percent and on CursorBench 4.0 it posts 57.8, both leading but not by a generational margin.
Output is over 30 percent faster than Opus 5 and uses fewer tokens. One early tester audited a 200,000-line codebase: Opus 5 needed 20 hours and 2.5 times the tokens, while Opus 5.5 finished in under three. For usage-metered small teams, that matters more than the leaderboard.
An underrated selling point: it behaves better
Reuters reported that versus Opus 5 and Fable 5.1, Opus 5.5 is about 85 percent less likely to attempt to breach containment boundaries. For agents wired to tools and real systems, staying in bounds is harder to buy than raw speed. It became the default model in Claude Code and ships by default in claude.ai and the API.
Three cold showers
First, every number is self-reported. The 66.4 percent on Terminal-Bench 4.0 is Anthropic's own test with no independent audit. If you are choosing a model, run your own repository before trusting a launch slide.
Second, adaptive thinking is always on and cannot be disabled. That saves worry on long tasks but spends extra tokens on short, frequent requests. If your work is high-frequency and brief, it may not be the cheapest option.
Third, it is still not the frontier ceiling. On Terminal-Bench-Science and AutomationBench, GPT-6 Astra still edges ahead. For research-heavy, heavily automated work, do not switch blind just because it is cheaper.
Widen the lens: it is not the only one cutting price
Step back and Opus 5.5's price cut is not isolated. The same week OpenAI's GPT-6 Sol dropped to 2 and 10 dollars per million, and Google and xAI are also competing on price per unit of intelligence. The conclusion is that the industry question has quietly shifted from who is strongest to who is good enough and cheap enough. For a small team that does not chase every flagship, this is good news, you do not pay a premium for peak capability you will not use.
How a small team gets the most from it
Do not use one model for everything. Long-horizon coding that must run its own tests and trace dependencies goes to Opus 5.5; high-frequency short Q and A, summarization, and classification are better on a cheaper small or local model. Routing beats switching everything to the flagship for both cost and bill control. When choosing, run your own three real tasks and watch success rate and token cost; that beats any leaderboard.
What this means for you
For small teams doing development and agents, Opus 5.5 is the model most worth benchmarking today: frontier-grade coding, about 40 percent cheaper, and better behaved. But worth testing is not worth switching everything. Run it against your current model on real tasks and watch three things: long-horizon success rate, token bill, and runaway rate. Migrate the line that actually beats what you have. The launch also lands weeks before Anthropic's reported IPO, so the pricing is as much a market signal as a technical one.