The cost of intelligence keeps falling. Open models are now close enough in performance that the choice matters less. I recently moved a large amount of repetitive inference to a server in my office. It saves me a lot of money.
The cost of intelligence keeps falling. Open models are now close enough in performance that the choice matters less. I recently moved a large amount of repetitive inference to a server in my office. It saves me a lot of money.
David Friedberg ran the list from the last ten days. Xiaomi's open-weight MiMo Pro scores near the frontier models. Bonsai 2 is a 27 billion parameter model, 5.9 GB, and it runs on a desktop. His line: you don't need data centers for 90% of what you can do with AI. Chamath Palihapitiya says the models are converging to within margin of error, so the edge moves to the harness you wrap around them.
The model is becoming the cheap part. If you want good AI, build a great harness.
Sources: David Friedberg and Chamath Palihapitiya, All-In, "Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails," Sep 25, 2026.