Anthropic released Claude Opus 4.1 this morning, an incremental but meaningful upgrade focused on agentic coding workflows. On SWE-bench Verified — a benchmark of real GitHub issues — Opus 4.1 scored 74.5%, narrowly edging out competing frontier models.
Computer use, in production
The model can now drive a virtual desktop: opening apps, filling forms, dragging files. Anthropic says the feature is stable enough to expose in the API behind a feature flag, with a recommended sandboxed VM.
- Vision + cursor coordinates returned in a single tool call
- Latency reduced ~40% versus the 4.0 preview
- Built-in safety classifier blocks risky actions like sending payments
💡 Tip: Pair Claude Opus 4.1 with the open-source `claude-agent-sdk` and you can ship a browsing copilot in an afternoon.
Pricing is unchanged from Opus 4: $15 / $75 per million tokens. A cheaper Haiku 4.1 follows next month.