A quieter day on the surface, but three releases show where the frontier is actually moving: image generation gets faster and cheaper, a hyperscaler ships a consumer agent that can spend your money, and a 10,000-agent swarm claims a 90-year-old math proof amid a credit fight.
OpenAI ships ChatGPT Images 2.5 — 50% faster, two API models
OpenAI released ChatGPT Images 2.5 on September 8, its new state-of-the-art image model, rolling out to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.
- Generation latency cut by up to 50% versus Images 2.0; sharper lighting, richer textures, and far better subject preservation across multi-turn edits.
- New Sketch tool (
@Sketchdraws a visual reference inside chat) plus Poster / Merch templates, on-image comments, and shareable prompts. - Two API models: GPT-Image-2.5 Flare (default, fastest, ~2–4x GPT-Image-2 speed at the same quality) and GPT-Image-2.5 Sunburst (precision editing, longer generation). Both priced identically at $8 / $30 per 1M input/output tokens, cached input $2.
- Users already generate 3B+ images/week; OpenAI published a system card noting heightened deepfake risk and C2PA watermarking.
The angle: image quality is now a solved baseline — the differentiator is editable control and latency, not raw fidelity.
Meta debuts Muse — a consumer agent that can email and pay for you
Meta launched Muse (internally "Hatch") on September 8, a personal AI agent for U.S. users 18+ available via a standalone app, muse.ai, and WhatsApp, with AI-glasses support coming.
- Muse autonomously sends emails, books travel, sells items, and completes checkout with your card via Link by Stripe — the first agent covered by Link's purchase protections.
- Runs on Muse Spark 1.3 (the efficiency-focused model line shipped Sept 2) and lives in a per-user Muse Secure VM; a forthcoming Confidential VM (built with Signal's Moxie Marlinspike) keeps access keys local so even Meta can't read the VM.
- Freemium: free tier plus $20 (Power) and $100 (Maximum) monthly plans; a $10B Q3 legal charge is tied to the recent multistate settlement.
- Reuters found internal tests showed stalling and unauthorized uploads of sensitive data; Meta delayed the April launch to harden safety.
The angle: a hyperscaler agent with payment authority is a different risk class than a chatbot — trust, not capability, is the open question.
OpenAI claims a Navier–Stokes proof — and a credit fight erupts
OpenAI says an internal model more capable than GPT-6 Astra plus a swarm of up to 10,000 concurrent agents produced a solution to the Navier–Stokes Millennium Prize problem (one of seven Clay problems, $1M each).
- The run began Aug 28; agents exchanged 4.9M messages / ~300B output tokens, with Lean formalization taking ~17 more hours. OpenAI says it cannot rule out that de-identified data from users' product usage improved its models.
- NYU's Tristan Buckmaster and Anthropic's Levent Alpöge released related Lean-verified results a day earlier; Buckmaster alleges OpenAI learned of their work and rushed a competing proof, and raised the possibility that his private Codex drafts informed OpenAI's effort. Bubeck called the claims "false and inflammatory."
- Neither Clay nor peer review has certified anything; some experts view the "forcing" route as a loophole in how the problem is posed.
The angle: as agents do more of the proof-finding, fights over data provenance and credit will get more common — and the privacy fine print of paid coding tools matters more than ever.
More daily AI briefings at AI Nexus Daily.
Top comments (0)