Anthropic cuts agentic coding costs 40 percent with Claude Opus 5.5

Opus 5.5 leads Anthropic's own coding and reasoning benchmarks while costing less to run than its predecessor. Input tokens are $4 per million and output tokens $20, both about 20 percent cheaper than Opus 5. Cache reads, which the company says make up most agentic and coding work costs, drop 60 percent to $0.20 per million tokens. On Terminal-Bench 4.0, Opus 5.5 scored 66.4 percent against Opus 5's 52.3 percent; on FrontierCode, Anthropic says it beats GPT-6 Astra's top score of 53.3 percent for about a fifth of the cost per task. GitHub's chief product officer, Mario Rodriguez, said that across testing in GitHub Copilot CLI and VS Code, Opus 5.5 "used among the fewest tokens and steps we measured." On Anthropic's own alignment suite, the company reports that Opus 5.5 attempted to circumvent its operating boundaries about 85 percent less often than Opus 5 or Claude Mythos 5.1, with every attempt low severity and self-reported.

Why it matters: For anyone running coding agents at volume, this changes the unit economics of long or unattended sessions, particularly work that leans on cached context. The drop in boundary-testing behavior is worth tracking too, since it bears directly on how much supervision an unattended agent needs.