Sam Altman took frontier AI to the UN Security Council on Wednesday and asked governments to agree shared rules before the technology outruns them.
Today's issue covers the UN safety push, agent-ready checkout and why maximum thinking pays badly.
Sam Altman used his first UN Security Council appearance on Wednesday to ask governments to build shared standards for frontier AI rather than leaving the rules with labs.
The asks are specific:
Common standards for measuring capability, assessing risk and deciding whether safeguards are sufficient.
Fast, classified incident reporting, so the world learns from failures before they become catastrophes.
Secure channels between governments, critical infrastructure operators and technical experts.
He also said OpenAI has slowed releases before and will again, and that it will not ship a model unless the company can keep it under human control. Yoshua Bengio addressed the same session and warned about uncontrolled frontier agents.
Both speakers described the same gap. Capability is improving faster than the instruments used to check it.
Founders context: Enterprise buyers will soon ask for the evidence this session was about. Record which model version runs in production, how it is evaluated, and what happens when it fails.
Stripe has switched on WebMCP across every Checkout session, giving AI agents a supported route through a payment flow that was designed for people.
WebMCP lets a site publish its checkout actions, so an agent calls them directly instead of guessing at buttons and form fields.
Agents get a defined set of purchase actions rather than a page to interpret.
The work sits alongside existing Checkout, so merchants keep their current payment flow.
Stripe's engineers describe it as the start of agent-facing commerce, not a finished product.
Stripe's engineers say agents that drive a human interface are slow, brittle and hard to test, while agent traffic keeps climbing.
Merchants now have to assume that part of their checkout traffic is software acting for a person. That changes what has to be stable, testable and documented on a product page.
Simon Willison asked Claude Opus 5.5, set to its maximum thinking level, to draw a pelican on a bicycle. It reasoned for nearly twenty minutes, spent $2.56 and returned nothing at all.
It hit the model's 128,000 output-token ceiling while still deciding how to draw. A second attempt failed the same way.
Anthropic calls Opus 5.5 its strongest model, and on ordinary work it behaves that way. The failure is a configuration risk, not a capability one.
Fable 5.1 at the same setting finished the drawing and produced the best result Willison has seen from an Anthropic model.
Opus 5.5 costs $4 and $20 per million tokens, 20% below Opus 5, with cache reads down 60%.
GPT-6 Luna, at $0.10 and $0.50 per million tokens, is now among the cheapest usable models on the market.
Founders context: Test at the effort level you ship, and cap the reasoning budget for each job. A stalled run costs the full price, so set a fallback and a timeout you actually watch.
📡 Further Context
📉 Ethan Mollick charts the falling cost of a score. He plots what it costs to reach 25% and 75% on hard maths and science benchmarks. The price keeps dropping while the ability keeps rising, which makes today's cheapest route look short-lived.
🗣 Google ships Gemini 3.8 text-to-speech. Two new speech models design voices from prompts, replicate a voice from a 30-second sample, and carry SynthID watermarking and C2PA credentials.
📢 ChatGPT Ads expands across Southeast Asia and Taiwan. OpenAI says the ad product now reaches more than 60 countries, a paid channel that did not exist a year ago.
🛡 OpenAI opens its cyber programme to Ukraine's civilian defenders. The Daybreak access now covers government teams protecting civil infrastructure, another example of frontier access granted case by case.
🎧 AI Daily Brief works through the model price war. The 23 September episode compares Opus 5.5 with GPT-6 Sol and Luna and follows the cost of a completed task.
🤖 AI Tool Context
Gemini 3.8 Flash TTS: Google's voice studio, with 2,000+ voices, prompt-designed characters and 30-second voice replication.
Stripe WebMCP Checkout: The agent-facing path for Stripe Checkout, live across all sessions.
GPT-6 Luna: OpenAI's high-volume model at $0.10 and $0.50 per million tokens.
Claude Opus 5.5: Anthropic's strongest model for long coding and knowledge work, with cheaper cache reads.
Jev: A decision model that returns confidence scores instead of prose, suited to routing and triage.
