Blog

Grok 4.6 ties GPT-5.6 Sol at a third of the price, and loses two benchmarks badly. I read xAI's own eval table row by row, then checked the number the launch post left out.

Grok 4.6 lists at $2 input and $6 output per million tokens. OpenRouter's measured effective input price is $0.74. Here is every meter on the bill, and which door you should buy through.

Grok Bot gives every AI teammate its own cloud computer and signs it into your real tools. Here is how it works, what it costs, and where the design breaks.

GPT-5.6-Cyber has no published per-token price, and OpenAI's own rate card has carried an empty Cyber row for generations. Here is what to budget with instead.

GPT-5.6-Cyber answers 95% of advanced cyber prompts where Sol answers 1.5%. What the model is, how Daybreak Red gating works, and where it loses to plain Sol.

Agent Plugins 1.0.0 shipped on 6 August 2026 with AWS, Cursor, Microsoft, OpenAI and Vercel behind it. Here is what it standardizes, and what it leaves out.

Anthropic kept Opus 5 at Opus 4.8 prices, but thinking is now on by default. Here is what a Claude Opus 5 run really costs once effort and caching are in.

A hands-on Claude Opus 5 review: what the benchmarks actually say, the hallucination rate that went up, and whether it belongs on a live support queue.

Fable 5 costs exactly double Opus 5. I went through both system cards, the docs and the independent benchmarks to work out when that second dollar buys anything.