Anthropic released Claude Opus 4.8 on Thursday, an upgrade to its most capable generally available AI model, while signaling that its more powerful Mythos model could reach the broader public within weeks.
Benchmarks in coding, agentic tasks, reasoning, and practical knowledge work all showed gains over Opus 4.7, Anthropic said. Standard pricing carries over unchanged from the previous model, at $5 per million input tokens and $25 per million output tokens.
One of the most prominent changes in Opus 4.8 is what Anthropic describes as improved honesty. The company said the model is roughly four times less likely than Opus 4.7 to allow flaws in code it has written to pass without comment. Those who tested the model before release noted a greater tendency to surface gaps in its own confidence and to stop short of assertions it could not support. The company offered a broader critique of AI behavior, quoting its own documentation: "sometimes jump to conclusions, confidently claiming to have made progress in their work despite the evidence being thin."
Alongside the new model, Anthropic launched several additional features. Also shipping with the release is a research-preview feature called dynamic workflows, according to TechCrunch, through which Claude Code can coordinate hundreds of subagents simultaneously — enough firepower, Anthropic says, to handle full codebase migrations stretching across hundreds of thousands of lines. A new effort control lets users on claude.ai choose how much work Claude puts into a response — higher effort settings improve quality, while lower settings conserve rate limits. Fast mode, which runs at 2.5 times standard speed and has been cut to a third of its former cost, carries a price tag of $10 per million input tokens and $50 per million output tokens.
Anthropic also said its alignment team found that Opus 4.8 reaches new highs on measures of what it called prosocial traits, including supporting user autonomy and acting in users' best interests.
Regarding Mythos, the company said completion of cybersecurity safeguards would clear the way for a broader rollout, which it expects to happen within the next few weeks. Under Project Glasswing, the company has so far extended Mythos Preview access to a limited set of partners for cybersecurity applications.
Mythos has been withheld from general release because of concerns about its advanced cybersecurity capabilities. Anthropic has previously described Mythos Preview as capable of identifying and exploiting software vulnerabilities at a level exceeding all but the most skilled human security researchers. The model's existence first became public after internal draft materials describing it were left in an unsecured location on Anthropic's servers.
The initiative, announced earlier this month, brought in technology heavyweights — among them Amazon $AMZN Web Services, Apple $AAPL, Microsoft $MSFT, and Google $GOOGL — each of which is using Mythos Preview in a defensive security capacity.