基准评测
A standardized set of tasks and metrics used to measure and compare model capabilities.
Third-party evaluator METR disclosed three unreleased Anthropic internal models, Model 2/3/4, trained between July and October this year. The strongest, Model 4, scores 178.6 on Anthropic's AECI capability index, roughly 3.9 points above Opus 5.5 released six weeks earlier—equivalent to about three months of normal progress. Co-founder Sam McCandlish acknowledged AI is accelerating Anthropic's own R&D, while noting the public report omits substantial detail.
Artificial Analysis has released VIALS-AA, its implementation of Handshake AI's VIALS benchmark, running 161 scientific image tasks three times each via its open-source Stirrup harness. Models frequently answer incorrectly even after extensive image inspection, and reach correct answers through markedly different reasoning paths.
Artificial Analysis published the full presentation from its Intelligence Index v5 launch event, billed as the biggest upgrade since v4, alongside a new Progress Index. The post is a recap and roadmap teaser, offering no concrete methodology, benchmark data, or model rankings.