Anthropic rolled out Claude Opus 4.8 today, billing it as a more effective collaborator for coding, reasoning and financial analysis. Early testers say the model judges agentic tasks sharper and flags uncertainty more often. The upgrade also claims a notable honesty boost, reducing unsupported claims.

Benchmarking shows Opus 4.8 achieving 69.2% on SWE‑Bench Pro, outpacing GPT‑5.5 and Gemini 3.1 Pro on that test, though GPT‑5.5 still leads on terminal‑coding. Alignment scores hit new highs for prosocial traits, with deception rates dropping below those of Opus 4.7 and matching the Claude Mythos preview.

Speed and cost improve dramatically: fast mode runs 2.5× quicker and the model costs three times less than its predecessor. New product features include Dynamic Workflows for Claude Code, letting hundreds of sub‑agents run in parallel on enterprise plans, and Effort Control in Claude.ai and Cowork, letting users trade response speed for quota usage.

Anthropic makes Opus 4.8 available across all plans with unchanged pricing, while promising cheaper models with similar capability and a more intelligent class beyond Opus. The company is also hardening safeguards for the Claude Mythos model, currently in limited preview, and expects broader rollout “in the coming weeks,” widening its enterprise AI toolkit.