OpenAI’s Dots walk into Muse’s house

The day split in two. Morning was about who sits between you and every click: Muse, Manus, Instinct, and an $8.2 billion AMD bet on the next camera view. Evening was OpenAI’s answer on stage — Simon Willison’s DevDay live blog watched Sam Altman introduce Dots, GPT-6.1 Sol, Ultrafast inference, a Decisions API, and a stack of Codex cloud tools. If you build agents for other people, the useful question is not which cute blob wins the App Store. It is which decisions you still refuse to hand the model.

Dots, Sol, and a sticky-note API

Dots look a lot like Muse on purpose: you name them, you get a blob avatar, voice does a lot of the work, and Astra powers the ambitious demos. ChatGPT Spaces are the shared-artifact layer. Dots already show up in Slack with their own identities. Pro and Enterprise get them today. Specialist dots for legal and finance are the enterprise pitch, including a Microsoft 365 thread.

The money slide for builders is GPT-6.1 Sol: near-Astra on coding and computer-use benches at about one-fifth the cost per task, listed at $2 input / $10 output per million tokens with cached input at $0.10. Ultrafast claims up to 300 tokens per second at six times standard price. Then the quiet product that matters for routing: a Decisions API on GPT-6 Luna that forces a pick from your predefined options in a fraction of a second — OpenAI’s reply to the small decision models that showed up this month.

Codex got the workbench treatment too: a full-screen CLI for parallel agents, git-worktree forks, and Security Cloud scans with Daybreak Blue on GitHub repos without a separate access form. Willison’s honest aside still stands — he had a bad Codex Cloud morning before the keynote — so treat the launch as a checklist, not a vibe.

The front door is still the product

Meta, Manus, and Instinct are still racing to own the layer above every shop and calendar. Muse gets a VM and a browser. Manus keeps cloud computers running after you leave. Cue gets phone, email, wallet, and computer. Instinct’s $1 billion Series C at a $10 billion valuation is the capital signal. Ben Thompson’s frame holds: when actions are abundant, the agent owns the relationship the way Google owned discovery.

Michael Spencer’s Muse piece puts numbers on the consumer side — Sensor Tower’s ~730,000 downloads in about five days after the 8 September launch — and warns retention looks soft. Amazon already blocked Muse from its store. That is the preview: personal agents only stay personal until a merchant decides they are unauthorized traffic.

Split the decisions. Keep the permission.

If DevDay was the product show, Prompt Engineering Institute’s decision-layer essay is the architecture you can steal tonight. Inside one “fix the bug” task an agent still decides which files matter, what failed, whether a command needs approval, and when the job is done. Giving every one of those to the same flagship model is convenient and expensive. The decision layer asks for the cheapest reliable mechanism given the cost of being wrong — code, a tiny classifier, a test runner, or a person — and keeps execution permission separate from the model’s opinion that an action looks fine.

That is the same lesson as Julia-1 (144M parameters, sticky-note options, no retrain when labels change) and as Jev beside Hermes: route skills and branches before the big model writes a novel. Zohre Shirazi’s workflow note is the human half — if AI clears the easy tickets, someone still owns the exceptions and the outcome.

They pulled a model. The IPO still warns.

OpenAI held Astra 6.1 after internal safety tests. The DevDay slot that morning looked empty; the evening keynote filled it with Dots and Sol instead. Anthropic’s leaked S-1 is already telling investors about catastrophic risk while booking hundreds of billions in compute commitments. Thariq’s Claude Code talk remains the engineering version of the same day: rewrite the harness, force implementation notes, and remember Exploit-Bench agents that used caches and GET-writable wikis as side channels.

AMD’s ~$8.2 billion World Labs deal is still the spatial swing — Atlas predicting the next view the way a language model predicts the next token. Watch for an eval you can rerun before you teach it.

I am watching whether Dots feel different from Muse once both have a wallet, whether Sol’s cached-token math holds up on a real overnight coding agent, and whether your decision layer starts as twenty sticky-note options or another system prompt. Until then, do not hand the intern your refund policy.

Also worth a click

• [[AINews] Opus 5.5 is good at explainer videos](https://latent.space/p/ainews-opus-55-is-good-at-explainer) — Latent.SpaceSimpleBench 88.4%; community video threads from about $4 to $1,874; Jev and Photon in the same dump.
• NVIDIA Kumo Tabular — Hugging Face / NVIDIAOpen tabular foundation model, 28M–215M, one-shot row labels, ranks first on four named benches.
• ProvenanceGuard for MCP agents — Hugging FaceSource-aware verification; blocked 138/139 expert-fail claims in the reported medical-agent test.
• How Yana Got to 1,000 Paid Subscribers — Solopreneur CodeConversion-first Substack playbook; 76% of paid upgrades in the first month.
• Jev-Omni multimodal classifier — AlphaSignal12B Gemma-tuned decision scorer across text/image/audio/video; independent of TypeSafe’s Jev.
• NVFP4 Qwen2.5-VL on Blackwell — AlphaSignalPre-quantized 7B VLM; claimed 2–3× vs FP8 throughput, ~3.5× less memory than BF16.
• DataTalks.Club ML Zoomcamp — AlphaSignal14,600 stars; 2026 cohort registration open.
• Gartner on vendor-built agentic AI — GartnerPredicts 70% abandonment of FDE-built agent products by 2028; thin press line.
• Frontend Info #32 — Frontend Info1,175-item dropdown 1,256 ms → 96 ms; modern CSS patterns.
• The Download: AI’s discovery problem — MIT Technology ReviewAnthropic enzyme-pattern claim vs biologists; plus the day’s safety/agent links.

New on arXiv

None named in today’s digest extras.