Claude Opus 4.8 favors reliability over a bigger model

Anthropic released Claude Opus 4.8 today, and the headline is not a bigger context window or a flashy demo. It is steadier work. The model scores 84% on Online-Mind2Web, a benchmark for browser agents, and Anthropic says it finishes every case end to end on the Super-Agent suite, where Opus 4.7 and GPT-5.5 stalled. The change that will matter most to people shipping code is quieter: Opus 4.8 is about four times less likely than 4.7 to read past a flaw in code without flagging it. Anthropic calls this honesty. For anyone using Claude to review pull requests, it is the difference between a reviewer who nods along and one who actually catches the bug.

Pricing stays at $5 per million input tokens and $25 per million output, the same as 4.7. Fast mode drops to $10/$50, three times cheaper than before, which makes the quick path far more usable. There is also a new effort control to trade quality against speed, and a Messages API change that lets you insert system instructions mid-task without breaking prompt caching. The release ships alongside dynamic workflows in Claude Code, where the model orchestrates hundreds of parallel subagents for jobs like large migrations. You can read the announcement on Anthropic's site. Anthropic also reports lower misalignment than 4.7, roughly on par with its Claude Mythos preview.

Why it matters

If you use Claude to review or write code, the drop in missed flaws is the change worth testing first, since a reviewer that quietly approves broken code is worse than no reviewer. The cheaper fast mode also changes the math for high-volume, latency-sensitive work.