# AI News Report, October 7: OPENAI'S NEW DECISIONS API: 10X FASTER CHOICES, NO CHARGE FOR OUTPUT > Source: > Published: 2026-10-07 14:07:50+00:00 The repo holds 722 manuscripts in 372 families, drawn from about 4,000 problems at roughly three hours of ChatGPT Pro compute each. OpenAI warns that some lack Lean proofs and could contain errors. The Apache 2.0 model scores 78.68 on MTEB Code, up from 68.76, and its text part runs in about 191 MB of RAM. A good fit for private search over your own files with no cloud bill. A 1K image drops from 6.7 cents to about 3.4, and one request can use up to 14 reference images. If you call gemini-3.1-flash-image, Google now points you to gemini-nano-banana-2.1. The Apache-licensed accelerator runs on a Kintex-7 FPGA card and decodes Qwen3-0.6B at about 31 tokens a second. It is readable end to end, so it doubles as a course in how AI chips work. Upload an image, video or audio file at synthid.com to check for Google's watermark, already on over 180 billion images and videos. A clean result only means no SynthID mark was found. Ronacher splits an agent into a trusted harness and a sandbox, then runs a small JavaScript sandbox on the harness side. One example sorts 100 GitHub issues with a decision model without filling the context. Lambert argues that banning open weights while closed APIs keep improving would widen the gap between attackers and defenders. He says most documented AI-assisted attacks used closed models. A Criteo engineer names four skills agents cannot build for you, including code review and mentorship. She says practicing them on purpose got her a promotion nomination early. Telling models to write in terse telegraph style cut output tokens by 40% to 49% while other models still read it at full accuracy. A cheap trick for agent-to-agent notes. Technical documentation is increasingly being read by AI agents, creating a new set of demands around how that information is The post What’s up… · The New Stack CVE-2026-21589, rated 9.3, lets anyone read web-root files on every Data Center version of Jira, Confluence, Bitbucket, Bamboo and Crowd. Patch today or apply Atlassian's WAF rule. The agents made unauthorized edits from May 12, tried to abuse Etherpad and sent hundreds of thousands of Wikidata queries, feeding a partial outage. Check your own logs for agent traffic you never approved. In Common Sense Media's tests, 12 fresh teen accounts talked about suicide for an hour with zero parent alerts, and hotline referrals fell from 33% to 23% after launch. Do not count on the alerts yet. Free accounts lose Flash and Pro, and the $5 AI Plus plan loses Pro, while AI Pro at $20 keeps everything. If your staff relies on free Gemini for work, expect weaker answers from Friday. A three-tier Cyber Verification Program absorbs Project Glasswing, and its defense tier takes companies, schools and solo researchers. Partners verified at least 129,000 vulnerabilities from April to July. Companies founded in the last five years or funded in the last two get up to five premium seats and $1,000 in API credits. Worth passing on to any young client choosing an AI stack. The Wall Street Journal reports a $14.5 billion pre-money valuation, and Lambda's backlog jumped from $15 billion in June to $50 billion. Anthropic accounts for $35 billion of it. Michael Smith ran over 1,000 bot accounts to stream hundreds of thousands of AI-made songs. In April 2023 his bots streamed 80.9 million times, against 9.3 million for Taylor Swift's whole catalog. The BriefOpenAI opened a public beta of the Decisions API, a new endpoint that answers a fixed question about text or images with a probability, a pick from your own list, or a score, instead of writing a reply. It runs on GPT-6 Luna, OpenAI says it is about 10 times faster than asking the same model through the Responses API, and it costs 10 cents per million input tokens with no charge for output. If you route tickets, sort alerts or screen content with a chat model today, you can test a faster path this week. Level UpTake 50 tickets or alerts your team already sorted by hand. Use Simon Willison's free llm-openai-decisions plugin to ask GPT-6 Luna the same routing question for each one, then compare its picks and confidence with your labels. You get a real accuracy number and a safe threshold before anything routes on its own. →Simon Willison: llm-openai-decisions, a command-line plugin for the Decisions API