Claude Opus 4.8 — Sharper judgement, dynamic workflows, effort control

Anthropic released Claude Opus 4.8 today, the newest version of its flagship model. It’s a modest but tangible improvement on Opus 4.7 across coding, long-running autonomous agents, and reasoning on complex enterprise problems — and at the same price.

The launch also brings effort control to claude.ai, a new “dynamic workflows” feature in Claude Code that can run hundreds of parallel subagents per session, and a cheaper fast mode.

What’s new

Claude Opus 4.8

The new flagship. Better at agentic tasks, sharper in judgement, and noticeably more honest — Anthropic’s evaluations show Opus 4.8 is around 4× less likely than Opus 4.7 to let flaws in its own code pass unremarked.

Pricing is unchanged from Opus 4.7:

• $5 per million input tokens
• $25 per million output tokens

Model ID: claude-opus-4-8. Available everywhere today via the Claude API.

Fast mode (research preview)

Fast mode for Opus 4.8 delivers the full Opus 4.8 intelligence at roughly 2.5× faster token generation. Pricing has dropped: it now costs only 2× standard Opus 4.8 pricing (was 3× on previous Opus releases) — $10 / $50 per MTok.

Effort control

A new control next to the model selector on claude.ai and Cowork lets users choose how much effort Claude puts into a response:

• Default is high effort — Anthropic’s best overall balance of quality and UX.
• Extra (“xhigh” in Claude Code) is recommended for difficult tasks and long-running async workflows.
• Max spends the most tokens for the best results.
• Lower effort settings respond faster and conserve rate limits.

Claude Code rate limits have been increased to accommodate the higher token usage at higher effort levels.

Dynamic workflows in Claude Code (research preview)

Claude can now plan the work and run hundreds of parallel subagents in a single session, then verify its outputs before reporting back. The marquee example: codebase-scale migrations across hundreds of thousands of lines of code, from kickoff to merge, with the existing test suite as the bar.

Available on Claude Code Enterprise, Team, and Max plans.

Mid-conversation system messages

The Messages API now accepts system entries inside the messages array. Developers can update Claude’s instructions mid-task — to change permissions, token budgets, or environment context as an agent runs — without breaking the prompt cache or routing the update through a user turn.

Self-hosted sandboxes for Managed Agents (public beta)

Build and deploy agents at scale while keeping sensitive files, packages, and services in your own infrastructure or with a managed sandbox provider.

MCP tunnels (research preview)

Connect agents to MCP servers inside your private network without exposing them to the public internet.

Honesty improvements

A general problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made progress despite thin evidence. Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims — borne out by Anthropic’s evaluations showing the ~4× drop in unflagged code flaws compared to Opus 4.7.

Anthropic’s Alignment team reports that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest,” with rates of misaligned behaviour substantially lower than 4.7 and similar to Claude Mythos Preview.

What’s next

Anthropic is working on lower-cost models that match much of Opus’s capability, and on a new class of higher-intelligence models codenamed Mythos. A small number of organisations are already using Claude Mythos Preview for cybersecurity work under Project Glasswing; broader release is gated on stronger cyber safeguards, expected in the coming weeks.

Try Opus 4.8

• API: claude-opus-4-8 — available everywhere today.
• AICoach model page: Claude Opus 4.8
• Anthropic announcement: anthropic.com/news/claude-opus-4-8

---

Source: Anthropic announcement, 28 May 2026.