Anthropic dropped two major AI updates on Thursday: the official release of Claude Opus 4.8 and a clear signal that the controversial Claude Mythos model will reach all customers “in the coming weeks.”
Opus 4.8: Faster, smarter, same price
Just six weeks after Opus 4.7, the new model posts meaningful benchmark gains while keeping the $5 per million input and $25 per million output token pricing. On the SWE-bench Pro coding test, it hit 69.2% (up from 64.3%), handily beating GPT-5.5’s 58.6% and Gemini 3.1 Pro’s 54.2%. It also leads on Humanity’s Last Exam (57.9% with tools) and OSWorld-Verified computer-use tasks (83.4%). The only setback: a second-place finish on the Terminal-Bench command-line suite. Anthropic also introduced “thinking effort” controls—Low, Medium, High, Extra, and Max—letting users dial how much compute the model spends. A fast mode runs 2.5× quicker but costs $10/$50 per million tokens.
Safety is a headline feature: Opus 4.8’s deception and misuse-cooperation rates are now “comparable to Claude Mythos Preview,” Anthropic’s most locked-down frontier system. It’s also four times less likely than 4.7 to let bugs in its own code go unflagged.
Mythos: Weeks away from wider access
Anthropic stated it expects to bring Mythos-class models to all customers shortly, the firm’s most explicit timeline yet. The model has been restricted to vetted partners through Project Glasswing after draft materials leaked in March described it as “the most powerful AI model we’ve ever developed.” Research from the UK’s AI Security Institute showed Mythos autonomously completing a 32-step simulated corporate network attack that normally takes human red teams 20 hours. Mozilla found 271 Firefox vulnerabilities, and security startup Calif used a preview to craft an exploit chain for Apple’s M5 chips.
Despite the safeguards, the announcement stirred prediction markets: on Myriad, the chance of a Mythos release by end of June jumped from 17.5% to 44% in a single day.
