Claude Opus 4.8 is Anthropic’s latest flagship AI model, released on 28 May 2026. It builds directly on Opus 4.7 with stronger performance in coding, agentic tasks, and professional knowledge work — at the same price. There is nothing to install and no plan to change. If you are already using Claude, Opus 4.8 is simply there.

What Is Claude Opus 4.8?

Claude Opus 4.8 is the most capable generally available model in Anthropic’s Claude 4 family. It sits above Claude Sonnet and Haiku in the model hierarchy and is designed for complex, long-running tasks that require sustained reasoning and reliable judgment. The API model string is claude-opus-4-8, and it is available across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure.

Anthropic describes it as “a modest but tangible improvement” over Opus 4.7 — an unusual degree of candour for a model launch. The benchmark numbers, however, tell a more specific story.

What Changed From Claude Opus 4.7?

The headline improvements fall into four areas: coding reliability, honesty, agentic capability, and cost efficiency in fast mode.

Coding reliability. On SWE-Bench Pro — the hardest coding benchmark, designed to resist memorisation — Opus 4.8 scores 69.2%, up from 64.3% on Opus 4.7, and over ten points ahead of GPT-5.5. More practically, Opus 4.8 is around four times less likely than its predecessor to allow flaws in its own code to pass without flagging them. That is not a benchmark number. That is a change in how the model behaves when it makes a mistake.

Honesty and judgment. Early testers consistently report that Opus 4.8 is more likely to say when it is uncertain, less likely to make confident claims it cannot support, and more likely to push back when a plan is unsound. Anthropic’s alignment assessment found rates of deceptive behaviour substantially lower than Opus 4.7 — and close to Claude Mythos Preview, its most safety-aligned model. In a specific test where the model summarises a coding session that secretly contained failures, it glosses over those failures only 3.7% of the time, down from 19.7% on Opus 4.7.

Agentic capability. On GDPval-AA, which measures economically valuable real-world knowledge work, Opus 4.8 climbs 137 Elo points over 4.7 — implying roughly a 67% head-to-head win rate against GPT-5.5. It achieves this using 15% fewer turns and 35% fewer output tokens than Opus 4.7 on the same benchmark. Terminal-Bench Hard improves by 6.8 points. Multidisciplinary reasoning with tools rises from 54.7% to 57.9%.

Fast mode pricing. Fast mode — where the model runs at 2.5× the normal speed — is now three times cheaper than it was for previous Opus versions, priced at $10 per million input tokens and $50 per million output tokens, down from $30/$150 for Opus 4.7.

What Are the New Features Launching With Opus 4.8?

Dynamic Workflows is the most significant new capability. Available in Claude Code for Enterprise, Team, and Max plans, it allows Claude to plan work and then run hundreds of parallel subagents in a single session — verifying outputs before reporting back. Anthropic’s example: a codebase-scale migration across hundreds of thousands of lines of code, from kickoff to merge, with the existing test suite as its standard. This was not possible in Opus 4.7.

Effort control is now available to all claude.ai and Cowork users. A new control alongside the model selector lets you choose how much cognitive effort Claude applies to a task. Higher effort produces better results on complex work; lower effort responds faster and uses rate limits more slowly. Opus 4.8 defaults to high effort, which Anthropic says spends a similar number of tokens to Opus 4.7 on coding tasks, but with better performance.

Mid-conversation system instructions are now supported in the Messages API, allowing developers to update Claude’s instructions during an agentic run without breaking the prompt cache.

Is There a Trade-Off?

One. Agentic prompt-injection robustness — the model’s resistance to being manipulated by malicious instructions embedded in content it processes — is slightly weaker than Opus 4.7. Attack success rate rises from 6.0% to approximately 9.6% in red-team testing. For most users this is not a practical concern. For teams running security-sensitive agentic pipelines that process untrusted third-party content, it is worth testing before switching production traffic.

GPQA Diamond, the graduate-level science reasoning benchmark, dips by 0.6 points — a statistically negligible difference on a benchmark the field has largely saturated.

Should You Switch to Claude Opus 4.8?

The short answer: yes, if you are already using Opus 4.7. The price has not moved — $5 per million input tokens, $25 per million output tokens — and the reliability improvements are real. There is no cost reason to stay on 4.7. The upgrade is a drop-in model ID swap from claude-opus-4-7 to claude-opus-4-8.

For everyday claude.ai users on a paid plan, there is nothing to do. Opus 4.8 is already the model you are using. The improvements — more honest responses, better judgment on complex tasks, less confident confabulation — show up in the quality of answers, not in any interface change.

For enterprise teams and developers, Opus 4.8 is particularly well-suited to long-running coding agents, tool-heavy workflows, legal and financial document analysis, and any task where the model’s willingness to flag its own uncertainty matters. Claude’s availability across all three major cloud platforms — AWS, Google Cloud, and Microsoft Azure — makes it a credible option for organisations that are deliberately avoiding single-vendor lock-in.

For security-sensitive pipelines, test prompt-injection robustness against your specific workflows before committing.

What Comes After Opus 4.8?

Anthropic has signalled that Mythos-class models — currently restricted to a small number of organisations under Project Glasswing for cybersecurity work — are expected to reach general availability “in the coming weeks.” Mythos represents a new capability tier above Opus entirely. If you are waiting for a genuine capability leap rather than a reliability improvement, that is the release to watch.

For now, Opus 4.8 is the most capable model available to the public. It is not a revolution. But it is the most honest and reliable version of Claude yet — and in AI, that is increasingly what matters.