Anthropic on Thursday released Claude Opus 4.8 and said its Mythos-class cyberoffense models, until now sequestered inside the partner-only Project Glasswing, will be available to all customers within weeks. “We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks,” the company said in a statement reported by TechCrunch.
The framing is its own news. Bloomberg has previously reported that Mythos was held back because Anthropic considered it too dangerous for the general public, and the model has lived inside Glasswing with Amazon, Microsoft, and Apple as approved partners, per Reuters. The same Glasswing post that announces the broader rollout still concedes that “no company — including Anthropic — has yet developed safeguards strong enough to prevent such models from being misused.” Both sentences are true at the same time, and Anthropic is publishing them on the same page.
What Mythos actually does is the unusual part. Anthropic says vulnerabilities surfaced by the Mythos Preview “have in some cases survived decades of human review and millions of automated security tests.” The company is also offering up to $100 million in usage credits for a “Claude Mythos Preview” tier priced at $25 per million input tokens and $125 per million output tokens, plus $4 million in donations to open-source security groups. Anthropic has been in ongoing discussions with US government officials about the model’s offensive and defensive use.
Opus 4.8 is the public-facing sibling and arrives 41 days after Opus 4.7, at the same pricing, positioned by Anthropic as its best publicly available coder. The San Francisco Examiner, republishing the New York Times, quotes Vals AI chief executive Rayan Krishnan saying the model scored 10% higher than its predecessor on the firm’s vibe-coding benchmark, and describes Mythos as “essentially a more powerful version of Opus 4.8.” Anthropic told Reuters the gains are “particularly in honesty,” with early testers reporting the model is “more likely to flag uncertainties around its work and less likely to make unsupported claims.” Bridgewater Associates, per TechCrunch, singled out exactly that flagging behavior as the upgrade’s biggest differentiator.
A research preview called Dynamic Workflows landed alongside it, designed to coordinate hundreds of parallel subagents, with Claude Code on Opus 4.8 pitched as capable of driving codebase migrations spanning hundreds of thousands of lines. The cadence matters: the prior Opus shipped to a chilly reception amid pressure from OpenAI’s Codex and Google’s Gemini Flash, and Anthropic’s current Sonnet is three months old, its Haiku seven. Opus 4.8 will be available through the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.
The structural read is straightforward. A lab that argued in March that a capability was too hazardous for open release is, in May, announcing the open release, with the original warning still posted verbatim on the same domain. The 2023 GPT-4 system card pioneered the genre of capability disclosure paired with restraint; the 2026 Glasswing post is its inversion, capability disclosure paired with monetization, and the safeguards arriving on the schedule of the product launch rather than the other way around.
Sources
- https://www.anthropic.com/glasswing
- https://www.bloomberg.com/news/articles/2026-05-28/anthropic-unveils-new-flagship-ai-model-that-s-better-at-coding
- https://money.usnews.com/investing/news/articles/2026-05-28/anthropic-to-roll-out-claude-mythos-in-coming-weeks-launches-opus-4-8
- https://techcrunch.com/2026/05/28/anthropic-releases-opus-4-8-with-new-dynamic-workflow-tool/
- https://www.sfexaminer.com/anthropic-unveils-new-more-powerful-claude-model/article_2d137abf-78b2-52cc-ad27-55f0bda15e0d.html