This Week AI Got Quieter and Cheaper (Oct 12, 2026)
The headlines were about bigger models. The real story was smaller, faster, and a lot cheaper to run.
The model-launch flood didn't stop this week, but the interesting moves were about cost, not size. OpenAI's GPT-6.1 Sol, out since September 29, lists at $2 per million input tokens and $10 out - about a fifth of GPT-6 Astra's rate - which matters more than any benchmark when you run it a million times a day.
Anthropic's Claude Sonnet 5.5, released September 30, doesn't cut its token price but reports 30 percent faster output and lower task cost in its own tests. For bug fixes and document work that compounds fast.
The stranger trend is decision models. TypeSafe's Jev kicked it off September 15 at $0.042 per million input tokens and no charge for output. By October 1, Perplexity, Cloudflare, and AWS had each shipped their own, most open-weight. These return a typed choice with a probability instead of prose, and they answer in milliseconds for cents.
What this means for anyone building: the routine yes/no steps in your agent - route, triage, approve - no longer need a chatty frontier model. Move them to a decision model and the bill drops by orders of magnitude. The open question is accuracy, and almost nobody has audited it yet.
I'd watch the decision-model layer more closely than the next flagship. It's where the actual money is being saved right now.