Claude Opus 4.8

Claude Opus 4.8 landed on 28 May 2026, only forty-one days after Opus 4.7, and closed out the Claude 4 generation before Opus 5 arrived in July. Anthropic pitched it as a more effective collaborator rather than a bigger brain: better agentic coding, multidisciplinary reasoning, computer use, knowledge work and financial analysis, at unchanged pricing of five and twenty-five dollars per million tokens. The numbers moved in two clusters. Coding rose to 88.6 percent on SWE-bench Verified and 69.2 percent on the harder SWE-bench Pro, with 1,890 on GDPval-AA — reported as 121 Elo points ahead of GPT-5.5. Reasoning moved further: mathematics on USAMO 2026 jumped from 69.3 to 96.7 percent, long-context tracking on GraphWalks at one million tokens nearly doubled from 40.3 to 68.1 percent, and multidisciplinary reasoning with tools went from 54.7 to 57.9 percent. Anthropic also reported the highest score it had recorded on its Legal Agent benchmark. Two product changes shipped alongside. Dynamic workflows, in research preview for Enterprise, Team and Max plans, let the model plan a large job and run hundreds of parallel subagents in a single session — aimed at work like whole-codebase migrations. Fast mode runs the same model at roughly two and a half times the speed for ten and fifty dollars per million tokens, which Anthropic described as three times cheaper than the fast option on earlier models. Effort control — high, extra high and max — became directly selectable in Claude.ai and Cowork. On behaviour, Anthropic reported new highs on its prosocial measures and substantially lower rates of misaligned behaviour than Opus 4.7; early testers described the model as more willing to flag its own uncertainty and less likely to assert things it could not support. The model stays available as claude-opus-4-8; wujec.ai keeps it as a separate profile from Opus 4.7 and Opus 5.

▸Videos

No videos yet.

▸Reviews