Introducing Claude Opus 4.8
Anthropic positions Opus 4.8 as a reliability-focused upgrade for coding and agentic work, alongside dynamic workflows, user-controlled effort, cheaper fast mode, and new mid-task instruction controls in the Messages API.
A model upgrade focused on judgment and reliability
Anthropic says Claude Opus 4.8 improves on Opus 4.7 across coding, agentic skills, reasoning, and professional knowledge work while keeping regular API pricing unchanged. Early testers highlighted better judgment, cleaner tool use, stronger long-session consistency, and a greater tendency to surface uncertainty rather than silently accepting weak assumptions.
Anthropic’s own evaluation says Opus 4.8 is around four times less likely than Opus 4.7 to let flaws in code it wrote pass without comment. The company also describes lower rates of misaligned behavior in its pre-deployment assessment, with full details in the model system card.
Dynamic workflows push Claude Code toward larger projects
The release arrives with dynamic workflows in research preview. Claude Code can plan a large task and run hundreds of parallel subagents in one session, then verify outputs before reporting back. Anthropic gives codebase-scale migrations across hundreds of thousands of lines as an example, with the existing test suite acting as an acceptance bar.
Effort control is also available in claude.ai and Cowork. Users can trade speed and rate-limit consumption for deeper reasoning, while Opus 4.8 defaults to high effort. Anthropic recommends higher settings for difficult or long-running asynchronous work.
API and pricing details matter for production teams
The Messages API now accepts system entries inside the messages array, allowing a harness to update permissions, token budgets, or environment context during a task without breaking prompt caching or pretending the update came from the user. Fast mode runs Opus 4.8 at up to 2.5 times normal speed and is priced at $10 per million input tokens and $50 per million output tokens; regular Opus 4.8 remains $5 and $25 respectively.
For agent builders, the release is notable because the model and the orchestration surface are evolving together. Better model judgment is useful, but dynamic parallelism, effort controls, cache-friendly instruction updates, and explicit verification are the pieces that determine whether that capability can be operated safely at scale.
This article is an editorial summary based on Anthropic News. For primary context and updates, read the original source. Anthropic News