Anthropic introduces Claude Opus 4.8

Anthropic introduces Claude Opus 4.8

Anthropic’s article, published May 28, 2026, announces Claude Opus 4.8 as an upgrade to Opus 4.7. Anthropic describes it as a modest but tangible improvement, citing benchmark results and reports from early testers who found it more reliable and better at judgment on agentic tasks. Anthropic says its evaluations found the model around four times less likely than its predecessor to leave flaws in its code unremarked; the article also reports lower rates of misaligned behavior than Opus 4.7. These are Anthropic’s reported evaluations and testers’ assessments, not evidence here of performance across all real-world uses. The release also brings effort controls in claude.ai and Cowork, and dynamic workflows in research preview for Claude Code. Anthropic says those workflows can run hundreds of parallel subagents and verify outputs; it gives codebase-scale migrations as an example. The article says regular-use pricing is unchanged from Opus 4.7, while fast mode is cheaper than for previous models. Anthropic’s stated next steps are to develop Opus-like capabilities at lower cost and strengthen safeguards for a more capable class of model; it says it expects Mythos-class models to reach all customers in the coming weeks. Whether the reported improvements carry over to users’ varied, consequential work remains unknown.

Publication: d0f31510-d34a-9e38-f9e9-5f3b501a7939

attribution — pass: This is a single Anthropic product announcement on its own news page. The reading consistently attributes the launch, the evaluation results, the tester reports and the plans to Anthropic or its early testers. It does not present them as established facts.

We’re upgrading Claude Opus to a new version: Claude Opus 4.8.

summary — pass: The reading leads with the actor and the change: Anthropic announces Opus 4.8 as a modest but tangible upgrade. It then covers the honesty and alignment results, the new features, pricing and stated next steps, and it keeps Anthropic's own modest framing. Benchmark, agentic and alignment are defined beside the reading. Subagents and Mythos-class are not explained, but the reading gives enough context (a more capable class of model) for a general reader. A tooltip for subagent could be added.

Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor.

date — pass: The reading states May 28, 2026, which matches the UTC day of evidence.originalPublished. The Mythos-class expectation is reported as Anthropic's forecast, not an outcome, and there is no hindsight framing. The page was modified on 2026-09-28. However, the cited passages are core launch-announcement content and plainly part of the original piece. Kind is publisher-page, so no archive disclosure is needed.

expect to be able to bring Mythos-class models to all our customers in the coming weeks

support — pass: Each claim matches the source text. These include the four-times-less-likely result, lower misalignment rates than Opus 4.7, hundreds of parallel subagents with output verification, unchanged regular pricing, cheaper fast mode, and the conditional Mythos expectation. The reading explicitly declines to turn evaluations into real-world reliability. The question follows from that gap. Although the page was edited later, the claims rest on integral launch text.

Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.

media — pass: No media candidates were supplied and none were selected. The proposed mediaId is null, which is consistent with the evidence.