A Chinese Model Beat GPT-5.5 for 1/10 the Cost — What's Real and What's Hype

The honest, no-hype brief on MiniMax M3 — the open-weights model a Chinese lab shipped on June 1, 2026 that beats GPT-5.5 and Gemini 3.1 Pro on a real coding test for about 1/10 the cost. What's genuinely real: on SWE-Bench Pro it scores 59.0% (ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2), it ships a 1M-token context window with native multimodal input, and its MiniMax Sparse Attention (MSA) keeps long-context compute cheap (~9x faster prefill / ~15x faster decode at 1M tokens). At a launch-promo $0.30/$1.20 per million tokens (standard $0.60/$2.40), it's roughly 5-10% the cost of leading US frontier models — the moat was a price tag, not a law of physics. But the viral reposts skip the asterisks: M3 is still behind Claude Opus 4.7 (64.3) on that same coding test, it scores under 12% on ARC-AGI-2 (weak abstract reasoning), every benchmark here is self-reported by MiniMax, and the 'open weights' are promised within ~10 days, not yet shipped. Includes the full benchmark scoreboard, the pricing tiers (and the >512K context cost cliff), a claim-vs-reality table, and an honest 'use M3 / use the best / wait' decision guide.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.