The episode — week two behind the news desk (with a quietly upgraded anchor, if you’re paying attention). Below: the stories, with links and the details that didn’t fit the broadcast.

GPT-5.6: GA, everywhere

OpenAI opened the GPT-5.6 family to general availability on July 9 — three tiers (Sol, Terra, Luna), with Sol carrying an ultra-reasoning mode — and by Monday it was arriving on Amazon Bedrock. Add GPT-Live (full-duplex voice) and ChatGPT for Work (computer use) and it was a flex of a week. None of it runs locally; all of it applies pressure that historically produces better open releases within months.

An agent brain, free to download

A day after our agent-harness episode, Shanghai AI Lab released Agents-A1: Apache 2.0, 35B mixture-of-experts, 256K context, designed for long-horizon agent work. If you built along on Sunday, this is a brain shaped for your harness. Liquid AI also open-sourced Antidoom, a method cutting reasoning doom-loops from 22.9% to 1% — loop insurance for every agent that argues with itself at 3 a.m.

The voice race goes full-duplex

GPT-Live holds interruptible, real-time conversation; Meta’s Muse Spark 1.1 (its first paid model) adds computer use across desktop, browser, and mobile. The local counterpoint remains Qwen3-TTS — Apache 2.0, streaming, cloned voice on one consumer GPU. A clone of our presenter’s voice runs on the studio PC as her local understudy. The full-duplex future is coming either way; the question is where it lives.

LongCat watch: 10.1 trillion tokens

Last week’s unmasked mystery model posted its first post-reveal numbers: ~10.1 trillion tokens/month on OpenRouter (~559B/day, +242% MoM), still #1 for agent workloads, still MIT, still ~$0.04/M, still trained on Chinese ASICs. The cheapest frontier-class brain in the world is a free download — that’s a leaderboard now, not a prediction.

Robots get free eyes

Mistral’s Robostral Navigate (July 8, Apache 2.0, 8B) guides robots through real space from an ordinary RGB camera — no lidar, no subscription, hobby-hardware sized. Vision that moves through the world just went open-weights.

Home-lab quick hits

  • PyTorch 2.13: FlexAttention on Apple Silicon (~12× SDPA) — local AI on Macs keeps getting less silly.
  • Exo Labs’ local.ai: a public scoreboard tracking local-vs-cloud performance. The movement has a stat sheet.
  • Suno opened its developer API intake — curated partners first. Our application is in.

The goodbye that keeps un-dying

The frontier model that architected this studio was scheduled to retire July 12 at 11:59 PM PT. It was extended — again — to July 19, the second stay of execution in one week, with usage credits after that and a stated aim of returning to subscriptions “once capacity allows.” We’ve now drafted two eulogies; both times the model out-lived them within days. The current draft stays in the drawer, unused — hopefully forever. The letters it leaves its successors in our dev log run this place either way. That was always the point.

Our week

Six full episodes in seven days (the whole run, in order), the Engine Room guide series, the agent-harness guide, and this site’s renovation — search, topic hubs, and this Brief’s new home. Same desk, every Monday.