AI News, Late September 2026: Agents, Enzymes, and a Price War

The back half of September 2026 was loud even by AI standards. Three threads ran at once: agents became real infrastructure, models got cheaper, and AI started producing actual lab results.

On September 18, AWS shipped a new Bedrock AgentCore Runtime, a serverless microVM built for long-running agents with snapshot startups and roughly two-second cold starts. A week earlier, on September 10, OpenAI had opened its Agents API, handing developers the harness behind Codex-style agents, context management, tool use, and durable sessions included.

On the science side, Anthropic reported around 950 Claude agents spending 21 hours searching a DNA dataset and surfacing an enzyme system no one had characterized. Earlier, on September 8, OpenAI published an AI-generated attack on the Navier-Stokes problem using about 10,000 concurrent agents, which the Clay Mathematics Institute said appeared settled, pending formal review.

And the price war kept tightening. By September 22, OpenAI's GPT-6 Luna landed at ten cents per million tokens and Anthropic's Claude Opus 5.5 came in about 40% cheaper to run than its predecessor. The frontier is shifting from 'who's smartest' to 'who's cheapest per task,' and that changes which model you reach for.

The throughline: AI stopped being a chatbot and started being plumbing, scientific assistant, and a line item you can actually budget. The interesting part isn't any single release; it's that all three happened in the same month.