Claude Opus 4.8: faster, more honest, and built for autonomous work

Anthropic released Claude Opus 4.8 today, May 28, 2026. It is the most capable model the company has made generally available. For Claude Cowork users this is a meaningful update: sharper judgment, more transparency about its own progress, and new infrastructure for running complex tasks autonomously.

What is new in Opus 4.8

The headline improvement is honesty. Anthropic describes Opus 4.8 as more likely to flag uncertainty about its own work and less likely to make unsupported claims. Early testers confirm that the model pushes back when it is not sure, rather than guessing confidently and moving on. That shift matters a lot in long autonomous workflows where a silent error compounds into a bigger problem downstream.

Alongside that, Anthropic is launching two new features alongside the model: a dynamic workflow system and an effort control panel.

Benchmark improvements

The numbers back up the quality claims. Opus 4.8 pushes the agentic coding score from 64.3 percent to 69.2 percent. Multidisciplinary reasoning with tools climbs from 54.7 percent to 57.9 percent. Anthropic also reports that Opus 4.8 outperforms competitors on financial analysis and broad knowledge work, two areas where Claude Cowork users lean on Claude the most.

Dynamic workflow: multiple subagents at once

This is the most structurally interesting addition. Dynamic workflow lets Claude spin up multiple subagents that run in parallel, coordinating them under one main thread. In practice that means a complex research task can fan out across several lines of inquiry at once, and a long coding project can split into components that build simultaneously.

For experienced Claude Cowork users who already run multi-step prompts, this formalises something you have been doing manually. Instead of chaining outputs one by one, Claude now manages the parallel execution itself. That is a genuine shift in what autonomous work looks like.

Effort control

The new effort control panel gives you a dial for how much compute Claude puts into each response. Low effort is fast and light for quick lookups. High effort gives you the deep reasoning pass for complex decisions. The same interface that runs dynamic workflows sits here, so you can mix and match: a high-effort planning phase followed by parallel subagent execution.

This is the feature that makes Opus 4.8 feel like the first model where the effort level is a first-class input, not just something implied by how you phrase the prompt.

Fast mode and pricing

Anthropic is also shipping a fast mode alongside Opus 4.8. It runs at 2.5 times the speed and costs three times less than the comparable output from previous models. Pricing between Opus 4.7 and 4.8 stays the same at the standard tier, so if you are already on a Pro, Max, Team, or Enterprise plan, the upgrade is automatic.

For developers the model is available on the Claude Platform natively and via Amazon Web Services, Google Cloud, and Microsoft Foundry.

What this means for Claude Cowork users

The combination of sharper honesty, dynamic workflow, effort control, and a fast mode changes the shape of what you can delegate. The model is more reliable about flagging when something is wrong. The parallel subagent architecture means multi-step projects move faster. And the effort control means you are not waiting for a deep reasoning pass when a quick answer is all you need.

The practical move today is to try a task you normally break into several sequential prompts and run it as a single dynamic workflow. The model handles the coordination. You handle the direction.

Opus 4.8 is available now on claude.com for Pro, Max, Team, and Enterprise users.