The frontier moat is evaporating.

MiniMax just hit 80.2% on SWE-bench with 10B active parameters. That's 20x cheaper than Opus 4.6 and nearly matching it on the hardest coding benchmark we have.

Meanwhile Google jumped ARC-AGI by 16 points in a single update and OpenAI is shipping 1000 tok/s models.

This week's ThursdAI covers all of it, with live interviews from both z.ai and MiniMax researchers, check it out