A flagship that got cheaper, not pricier
Anthropic released Claude Opus 5.5 on September 22, the newest version of its flagship large language model for coding, reasoning, and long-running agentic tasks. The headline is unusual for a flagship release: it costs less than the model it replaces. Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens, a 20% cut from Claude Opus 5's pricing, with cached-token reads priced at $0.20 per million. The model also carries a 1 million token context window with no premium pricing for using the full window, and supports responses of up to 128,000 tokens.
What improved
Anthropic and early testers point to gains concentrated in coding and multi-step work rather than a broad, even lift across every task type. On Terminal-Bench 4.0, a benchmark that scores how well a model handles real command-line and agentic workflows, Opus 5.5 scored 66.4%, up from 52.3% for Opus 5 and ahead of Anthropic's own mid-tier Claude Fable 5.1, which scored 55.8%. On FrontierCode v1.1, a coding benchmark, it scored 54.4% against 48.0% for Opus 5. On CursorBench 4.0 it reached 57.8%, against 46.6% for Opus 5. For knowledge work rather than coding, Anthropic's GDPval-AA v2.1 evaluation put Opus 5.5 at an Elo rating of 1846, ahead of Fable 5.1's 1735 and Opus 5's 1708, according to figures reported by VentureBeat. Anthropic's own benchmark table also lists an 89.9% score on SWE-bench Pro, a test that grades a model on real, unresolved issues pulled from software repositories rather than curated coding puzzles, which the company presents as its strongest published result for the release.
GitHub's chief product officer, Mario Rodriguez, said the model "used among the fewest tokens and steps GitHub measured" in its own testing, and that it solved more tasks in fewer than half the steps Opus 5 needed for the same work. Anthropic also cites early use by Lovable, Spotify, and Optiver, though those are companies Anthropic selected to highlight rather than independent benchmarks.
Anthropic's own safety numbers
Alongside the performance claims, Anthropic published results from what it describes as a roughly 2,000-scenario automated behavioral audit, reporting that Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior the audit tracks. In a separate containment evaluation designed to test whether the model tries to work around imposed restrictions, Anthropic says Opus 5.5 attempted to circumvent boundaries roughly 85% less often than either Opus 5 or Claude Mythos 5.1. The model was also tested before release by outside evaluators, including the AI safety groups Frontier Design and METR.
The competitive backdrop
Opus 5.5 arrives into a market where rivals have been cutting prices as aggressively as they add capability. OpenAI cut prices roughly in half with its GPT-6 Sol and Luna models earlier this month, and Alibaba has been pushing its own frontier ambitions, unveiling a new AI chip on September 22 alongside plans for a model with between 5 trillion and 10 trillion parameters. Anthropic's own leadership has been vocal on the other side of that race: chief executive Dario Amodei has publicly called for the industry to slow the pace of frontier AI development, a position several rivals have since echoed. Opus 5.5's pairing of a price cut with an emphasis on safety-audit results reads as an attempt to compete on both fronts at once, capability and caution, rather than choosing between them.
What it means for developers
For teams already building on Claude, Opus 5.5 is a straightforward upgrade path: same API, lower per-token cost, and a context window large enough to hold entire codebases or long document sets without chunking. The bigger practical change is the token and step efficiency GitHub reported, since agentic workflows that call a model repeatedly in a loop are often bottlenecked as much by how many round trips a task takes as by the price of any single call. Teams currently paying for Opus 5 should benchmark their own workloads against 5.5 before assuming the published benchmark gains translate directly, since Anthropic's own figures show the improvement is uneven across task types, largest in agentic coding and smaller elsewhere.
