What is Claude Opus 4.8, and what is new in it?

Yesterday, Thursday, May 28, 2026, Anthropic released Claude Opus 4.8, the new version of its most capable widely available model. It replaces Opus 4.7, which came out in April, and keeps the same price. Anthropic itself describes the upgrade as "a modest but tangible improvement on its predecessor," an unusually restrained phrase for a launch.

The model is only part of the news. Anthropic also gave everyday users a way to choose how hard Claude works on each answer, gave Claude Code a way to split huge jobs among hundreds of helpers, and cut the price of its faster mode to a third.

What Anthropic released

Opus 4.8 is available on claude.ai, in Claude Code (Anthropic's coding agent) and in Cowork (its assistant for office tasks), and to developers through the Claude API under the name claude-opus-4-8.

The scores come from Anthropic's own chart, as reported by The Decoder. Humanity's Last Exam is a set of very hard expert questions across many subjects. A token is a small chunk of text, often part of a word; you pay for the tokens you send and the tokens the model writes.

There is a cost wrinkle worth knowing. The Decoder notes that Opus 4.7, though priced the same as its predecessor, cost 30 to 40 percent more in practice because it used more tokens. The independent testing firm Artificial Analysis found that Opus 4.8 writes about a third fewer output tokens (35 percent) than its predecessor on a set of office-style tasks, so real bills may fall even though the price list has not changed.

The headline change: more honest about its own work

Anthropic puts honesty at the top of its list. A common weakness of AI models is jumping to conclusions: reporting that a task is done, or that code works, when the evidence is thin. Early testers say Opus 4.8 more often points out what it is unsure of and makes fewer claims it cannot back up.

Anthropic measured this in coding. Opus 4.8 is about four times less likely than Opus 4.7 to leave a flaw in code it wrote without mentioning it. For anyone who hands an AI a task and walks away, that matters more than a few benchmark points: a model that says "I could not get the tests to pass" is far more useful than one that says "done."

Anthropic also ran its usual alignment assessment, a set of tests for whether the model behaves as intended, for example avoiding deception and refusing to help with misuse. It reports that such misbehavior is substantially lower than in Opus 4.7 and similar to Claude Mythos Preview, the more powerful model it currently shares only with a small group of security organizations.

Effort control, dynamic workflows and a cheaper fast mode

Effort control. A new setting next to the model picker in claude.ai and Cowork lets you choose how much effort Claude spends on a reply. Higher effort means Claude thinks more often and more deeply before answering; lower effort gives faster replies that use up your plan's usage limits more slowly. It is available on all plans. The default for Opus 4.8 is high.

Dynamic workflows. In Claude Code, a research preview for Enterprise, Team and Max plans lets Claude plan a large job and then run hundreds of subagents in parallel. A subagent is a separate copy of Claude handed one slice of the work, like a team where each person takes a chapter. Claude then checks the combined result before reporting. Anthropic's example is moving an entire codebase of hundreds of thousands of lines to a new setup, from start to merge, with the project's existing tests as the standard to meet.

Fast mode. Opus 4.8 can run at 2.5 times its normal speed for $10 and $50 per million input and output tokens, a third of the earlier fast-mode price.

For developers, the API now also lets a program update Claude's instructions partway through a task, for example to change its permissions while an agent is running.

Why it matters

Opus 4.8 arrives five weeks after GPT-5.5 and nine days after Google's Gemini 3.5 Flash, and Anthropic's chart puts it ahead of GPT-5.5 on most of the tests it highlights. The bigger signal is where Anthropic says it is heading. It plans cheaper models with much of Opus's ability, and it expects to offer Mythos-class models, a new tier above Opus, to every customer within weeks, as soon as stronger cybersecurity safeguards are ready.

For users, the effort control is the most practical change: you can now decide when speed matters more than depth, instead of the model deciding for you.

Which effort level should you choose in Claude?

Start with the default, high. Anthropic picked it as the best overall balance, and for coding it uses about as many tokens as Opus 4.7's default while performing better.

Drop to a lower setting for simple, frequent requests: rewording an email, summarizing a short page or answering a quick fact question. You will get faster answers and stretch your usage limits further.

Raise it to extra for hard reasoning, long documents or jobs you will leave running, which is what Anthropic recommends for difficult and long-running tasks. Keep max for the few problems where you would gladly wait longer for the best answer. A simple habit helps:

Before you start, tell me in one line whether this task needs
more effort than usual and why. Then do the task, and end with
a short list of anything you are not sure about.