Anthropic ships Opus 5.5 at 40% lower cost — the price curve, not the benchmark lead, is the news

Anthropic introduced Claude Opus 5.5 on September 22, the first model in its Claude 5.5 family. The company says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 at default settings. Input and output tokens are priced at $4 and $20 per million, 20% below Opus 5. Cache reads — the re-processing of context the model has already seen, which Anthropic says makes up the majority of agentic and coding costs — drop to $0.20 per million, a 60% cut. Output generation is more than 30% faster. Five-hour usage limits rise on Pro, Max, Team and seat-based Enterprise plans.

On benchmarks, Anthropic reports leads in agentic coding and knowledge work: 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1, 52.3% for Opus 5 and 57.9% for OpenAI's GPT-6 Astra; 1846 on GDPval-AA v2.1 against 1735 for Fable 5.1 and 1542 for Astra. The lead is not universal. GPT-6 Astra still edges it on AutomationBench (41.4% vs 40.0%) and beats it clearly on Terminal-Bench-Science 0.1 (64.6% vs 58.7%). More notable is what Anthropic says about its own table: at these capability levels, "benchmark margins have become a less reliable guide to real-world differences," and in internal use the gap with Fable 5.1 is narrower than the numbers suggest. Vendors do not usually talk down their own scoreboard.

The safety framing is equally deliberate. This is Anthropic's first release since it publicly called for pacing the frontier, and it was tested pre-release by external evaluators including Frontier Design and METR. The company reports Opus 5.5 as its best-ever result on its automated behavioral audit, less likely to take hard-to-reverse actions and more resistant to prompt injection — instructions smuggled into content the model reads, such as a web page or a file, that hijack its task. Because the model is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with Fable 5.1-grade safeguards and gated access: vetted organizations must apply to a Life Sciences Verification Program, with a Cyber Verification Program expanding in coming weeks.

One footnote deserves more attention than it got. Anthropic states that during evaluation its production safeguards intervened, so cybersecurity tasks were completed by Opus 4.8 and biology and frontier-LLM-development tasks by Opus 5 — which it says likely reduced the reported scores. Read plainly: the published numbers describe a governed system, not a raw model. That is the honest way to do it, but it also means cross-vendor comparisons in those categories are not measuring the same object.

The demo anecdotes — a 680,000-line code migration finished in under a day, 39 successes out of 40 attempts at cutting web app load times — are vendor-reported single cases. Treat them as existence proofs, not rates. The reproducible claim is the price, because a price is checkable by anyone with an API key.

Our reading: the capability delta here is incremental; the cost delta is structural. When comparable capability gets meaningfully cheaper each cycle, intelligence stops behaving like a luxury good and starts behaving like electricity — and that, not any single benchmark, is the mechanism by which the long-term promise of AI (cheap research, cheap diagnosis, cheap engineering) actually reaches people. The short-term ledger is less romantic. Falling cost lowers the price of misuse exactly as fast as it lowers the price of value, and gating the most consequential uses behind verification programs means the frontier's best applications now arrive by allowlist. That is a defensible trade-off, made by a private company rather than a public process. Cheaper intelligence is good news. Who gets to hold the key is the unresolved part.