Model evaluations and how to read them.
4 posts
Cerebras's CS-4 puts three WSE-3 wafers, clocked twice as fast, in one rack. I measured its 30x claim: the GPU behind it does about 180 tokens a second.
xAI, Anthropic and OpenAI shipped Grok 4.7, Claude Opus 5.5 and GPT-6 Sol and Luna in two days. Prices, capabilities, caveats and how to choose.
ARC Prize says OpenAI's GPT-6 Astra scores 62.7% on ARC-AGI-3 with its standard harness and 99.9% with OpenAI's own memory settings. What the gap means.
Anthropic's Fable 5.1 ships with 75% cheaper cache reads, a safeguards layer your code must handle, and API changes that break history-editing harnesses.
