AI benchmark analysis and model verdicts

A single feed of every model verdict and every explainer we've published, in the order we published them.

Verdicts are head-to-head calls on a specific launch chart — DeepSeek V4 Pro vs. Claude Opus 4.8 is the first — read against the same real/tie/unverified rules as every table on this site. Blog posts step back from any one matchup to explain a pattern across the data, like how often vendor-reported scores drift from independent reruns. Both are held to the same standard: no number printed without a source, no claim that outruns what the source actually shows.

• Analysis2026-08-28

  Models with 10M token context windows 2026

  None of the 35 models we track ships a native 10M-token window. 24 run flat-rate to their max; 10 price-step at or above a threshold. Every claim sourced.
• Analysis2026-08-27

  GLM-5.3-Flash vs DeepSeek Flash vs Qwen3.8

  GLM-5.3-Flash vs Qwen3.8-Flash-Next vs DeepSeek V4 Flash: independent-score counts ran 4/6, 0/4, 8/8 at launch — two days later the boards moved.
• Analysis2026-08-24

  GPQA Diamond leaderboard 2026

  We graded our own GPQA Diamond leaderboard a D and stopped ranking it. Here is the data behind that call, and what changed across all 253 model pairs.
• Verdict2026-08-17

  DeepSeek V4 Pro benchmarks: real or noise?

  Its launch chart shows it beating Claude Opus 4.8. About half of that chart deserves your attention.
• Blog2026-08-17

  Self-reported vs. independent AI benchmarks

  When we counted, 29 of our 42 scores were vendor-reported. By day's end we'd swapped most for independent runs — one vendor claim was 5.6 points high.