# AI News Report, September 30: OPENAI: GPT-6.1 SOL MATCHES ASTRA ON CODING AT A FIFTH OF THE PRICE > Source: > Published: 2026-09-30 14:15:25+00:00 Dots run on GPT-6 Astra, answer in ChatGPT, Slack or Teams, and reach Pro and Business Premium first. Specialist dots with their own company identity will plug into Microsoft Agent 365. It keeps checking new commits and filters duplicate findings, on Pro, Business, Enterprise and Edu plans. A new Ultrafast tier runs up to six times faster in the API at six times the price. Google Research reports gains of up to 14.1 points on training tasks and 4.7 on unseen ones, with 30 percent fewer tokens. The code is on GitHub, so you can try it on your own agent. Cloudflare built it to treat agents as customers across 3,500-plus API operations, and released it under Apache 2.0. One spec can now feed every surface an agent reads. It runs 78 hard questions daily for 30 days against an Opus 5 control, and after 6 days it has found no change. Borrow the method if you think your own model got nerfed. He flags Ultrafast at up to 300 tokens a second for six times the price, and says developers have wanted Sign in with ChatGPT for years. The quickest way to see what shipped. He runs TypeSafe AI's Jev on 25,000 IMDb reviews and gets 96.47 percent accuracy for about 65 cents. A cheap way to sort tickets or route requests without a big model. The Pi team changed its mind because MCP got better, and now runs tool calls through a harness-side JavaScript layer called Codemode. Worth reading if you are deciding whether your agent needs MCP. His great-great-grandfather was a village farrier who became a mechanic when cars arrived. His advice to developers worried about AI: keep the why, let go of the how. Six months ago I wrote about how to kill the code review, and I was wrong. I still think line-by-line The post Kill the code review theater, keep the… · The New Stack Each lab commits to internal controls, an oversight team, outside auditors and an independent review committee. A separate order tells agencies to say "super intelligence" instead of AI. For now it points people to other websites, and features like passport applications are listed for 2027. The budget office has 90 days to tell agencies how to hook in. Anthropic opened an incident at 14:21 UTC Tuesday and applied a fix by 14:41, though some sign-ins and sessions kept failing. Have a fallback model ready for work that cannot wait. Only 104 of 186 large Dutch data centers reported to the national agency, despite an EU rule for sites of 500 kW and up. Dutch data centers used 4.6 percent of the country's power in 2024. TVA's 300 MW BWRX-300 at Clinch River in Oak Ridge, Tennessee, got its permit about 17 months after applying. The first one sets the pace for the rest. Researchers recovered 2,700 directory inodes from other workloads, and Cloudflare Sandboxes, built on Containers, were hit too. It was fixed the day it was reported, and Cloudflare saw no sign of outside use. The BriefOpenAI released GPT-6.1 Sol at DevDay on Tuesday and says it matches GPT-6 Astra, its top model, on a hard coding test at about a fifth of the cost (tokens are the chunks of text AI models bill by). It lists at $2 per million input tokens and $10 per million output, the same as GPT-6 Sol, and cached input drops to 10 cents. If you pay for GPT-6 Sol or Astra through the API, test gpt-6.1-sol on your own work this week, and read its safety card first: OpenAI rates it Critical for cybersecurity. Level UpBefore you switch models, run 20 of your own real tasks through the old model and the new one and compare the answers side by side. Then read the GPT-6.1 Sol system card addendum, which shows where it does worse, including a coding misrepresentation rate of 1.50 percent against 1.30 percent for GPT-6 Sol. →OpenAI: GPT-6.1 Sol system card addendum