Anthropic launches Claude Opus 4.8

Anthropic has launched Claude Opus 4.8. The new model replaces Opus 4.7 as the most powerful Opus version and, according to Anthropic, is better at coding, agent tasks, reasoning, and professional knowledge work. The price for standard use remains the same.

According to Anthropic, the most significant improvement lies in reliability and collaboration. Opus 4.8 is said to handle uncertainty better, detect its own errors more frequently, and be less likely to make unsubstantiated claims that a task has been completed successfully. Anthropic states that in evaluations, the model lets code errors go undetected approximately four times less often than Opus 4.7.

At the same time, Anthropic is launching new features related to agentic work. Claude Code is getting “dynamic workflows,” which allow Claude to schedule large tasks, start hundreds of sub-agents in parallel, and verify results before reporting back. Anthropic cites codebase migrations involving hundreds of thousands of lines of code as an example.

Users of claude.ai and Claude Cowork will also receive effort control. This allows them to choose how much thinking Claude puts into a task. Lower effort yields faster responses and saves rate limits; higher effort is intended for difficult or long-running tasks. In Claude Code, the highest setting is called xhigh, among other things.

The API update is also relevant for developers: the Messages API now accepts system entries within the messages array. As a result, an agent can update instructions, permissions, or context during a task without breaking prompt caching or having everything run via a user turn.

Opus 4.8 is available starting today. Regular usage costs the same as Opus 4.7: $5 per million input tokens and $25 per million output tokens. Fast mode is faster and, according to Anthropic, three times cheaper than with previous models, at $10 per million input tokens and $50 per million output tokens.

For AI users and businesses, this release shows that the battle between frontier models increasingly revolves around agent reliability. Not only smarter answers, but better planning, correction, reporting doubts, and completing long workflows is becoming the core of the competition.