Models, gated and open: the AI release roundup
A dated roundup of the AI launches that matter this week, minus the marketing.
Three things moved in AI this week. I'm skipping the hype and giving you the dated facts, because the dates are the only part that matters.
Gemini 4 Argon went out to cyber defenders only (Sept 30, 2026). Google's flagship scored 53 on Artificial Analysis's index, level with GPT-6 Astra, at an introductory $2/$10 per million tokens. But it is gated to vetted cyber defenders through the Fairwind Program, so most teams can read about it and not use it yet.
Mistral opened Large 4 in preview (Oct 6, 2026). A one-trillion-parameter model at $1.36/$4.18 per million tokens, scoring 38 on the index, with open weights promised by month's end. It is the European lab's clearest shot at the front row.
ElevenLabs and image models moved too (Sept 8 - Oct 6, 2026). Eleven v4 and v4 Turbo narrowed the gap to real voiceover. Google's Nano Banana 2.1 (Oct 6) and OpenAI's GPT-Image-2.5 (Sep 8) pushed editable image generation forward. DeepSeek-V4.1-Flash (Sep 10) kept the open-weight pressure on.
The pattern: the flashy agents get the headlines, but the gated flagships and the cheap decision models are what will actually change how we build. Watch those.