What's new in DevEx - Septem

OpenAI launched GPT-6 Astra at $10 and $50 per million tokens, a 2.5x increase over GPT-5.6 Sol. Greg Brockman said "Welcome to the AGI era." The headline 99.9% on ARC-AGI-3 came from a Provider Adapter harness, and ARC Prize published the standard harness result at 62.7% for the same model. Artificial Analysis measured the Intelligence Index at 61, unchanged from Sol and behind Claude Fable 5.1 at 66. Requests above 272,000 input tokens pay double.

Silicon Data's LLM Token Expenditure Index fell to 97 cents per million tokens on September 1, less than half its summer high. CNBC credits Chinese open-weight models, OpenAI price cuts, and the spread of dynamic pricing. The same week Alibaba's Qwen3.8-Max-0902 debuted at number one on Code Arena WebDev with 1,691 points, three above Claude Opus 5, at a blended $5 per million. Google's Gemini 3.8 Flash landed at $0.75 input. The floor keeps dropping while frontier list prices go up.

Claude Fable 5.1 launched alongside Mythos 5.1, the same model with minimal safeguards, invite-only through Project Glasswing. Terminal-Bench-Science more than doubled to 52.6%, and Artificial Analysis put the Intelligence Index at 66, the highest it has measured and five points above GPT-6 Astra. The twins differ by five points on Terminal-Bench 4.0, 60.9% to 55.8%, which is a direct measurement of what refusals cost. Claude Code users see roughly 60% fewer cyber-safeguard interruptions per session. Cache reads dropped 75% to $0.25 per million, and three breaking API changes ship with it.

OpenAI announced that it is winding down the contract supplying models to Cursor, with a proposed shutoff of November 12. It cited SpaceX and xAI terms-of-service violations and its own responsibilities around the Astra release. SpaceX closed its $60 billion acquisition of Anysphere on June 16. Cursor did nothing to trigger this, its owner did. Every product whose core value runs on a competitor's API now has a dated example of how that ends.

Cursor shipped self-hosted machines for Cloud Agents, with backends across AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B. Codebase, build outputs, and secrets stay on your infrastructure, and self-hosted workers support computer use on Linux and Mac. Vercel shipped the matching half a day later, running Cursor agents in Firecracker microVMs with scale-to-zero workers. DoorDash is pushing 130,000 engineering tasks a month through isolated microVM agents. The question moved from whether agents can run code safely to which pool you point them at.

PostHog just added nineteen MCP tools covering error-tracking rules, visual review, customer analytics, integrations, data pipelines, warehouse sources, and signals. The self-driving inbox was rebuilt around a three-tab layout with right-click actions that open pull requests directly, and scouts can now DM reports to five people in Slack instead of requiring a channel. A follow-up batch added asyncio-native event capture for Python and Next.js route templates in pageviews.

Till next time,