May 29th news, this morningAnthropic Official launch of the flagship model Claude Opus 4.8. The ability to program, AI intelligent body tasks, reasoning and knowledge work has been upgraded, while maintaining the same pricing。
standard model pricing maintains token 5 dollars per million input, token 25 dollars per million output; fast-track model pricing is token 10 dollars per million input, token 50 dollars per million output, which is about 2.5 times faster than in the previous generation and reduces model costs to one third of the previous generation。
At the heart of this upgrade is judgement and honesty. Official assessments indicate that the probability of allowing the self-code defects to pass without explanation is about a quarter of the previous generation of Opus 4.7, preferring to proactively mark uncertainty and reduce the lack of evidence-based conclusions。
For baseline tests, Opus 4.8 scored 69.2% on SWE-Bench Pro, above 64.3% in Opus 4.7, and exceeded GPT-5.5 and Gemini 3.1 Pro in a number of tests, such as Agentic Coding, Agentic Computer Use, Knowledge Work, Finance Agent. However, on the terminal programming benchmark Terminal-Bench 2.1, the GPT-5/782% was still above the Opus 4.8 74.6%。
Anthropic also indicated that a more capable Claude Mythos series model is planned for all clients in the coming weeks。