The model now hears, thinks, then answers.

OpenAI shipped new realtime voice models in the API — reasoning, translation, and transcription rolled into the same realtime stream. The headline isn't latency. It's that the agent now decides, mid-utterance, whether you're asking a question or thinking out loud.

Five-hour limits, doubled — and a Mars-side dependency.

Anthropic doubled Claude's 5-hour usage windows on Pro, Max, Team, and Enterprise, removed peak-hour throttling, and bumped Opus rate limits — paid for with a compute deal that taps Colossus, the SpaceX-allied datacenter cluster. The story is partly about ergonomics, mostly about who supplies the next rack.

Copy Fail: a kernel hole, patched at fleet scale.

When a Linux privilege-escalation vulnerability — call it Copy Fail — landed in the kernel community, the response was the headline: detected, mitigated, and patched across Cloudflare's global fleet before exploit code circulated. No customer impact, no traces of malicious activity, no scheduled emergency window.

Forty-four libraries, no more monkey patches.

Sentry is shipping native TracingChannel hooks into 44 JavaScript libraries — Express, Next, Fastify, Apollo, Prisma, and the rest. Observability moves to the library boundary, where it belongs. Drop-in agents stop fighting bundlers, source maps, and ESM hoisting; spans match the code authors meant.

The prior-restraint era begins.

By Zvi Mowshowitz · AI #167

For a decade, frontier labs trained models, then released them when they wanted. Zvi argues that arrangement is over. Increasingly, model releases require an affirmative permission — from one or another arm of government — before weights or APIs leave the building. The mechanics are uneven, the precedent is set, and the question is no longer whether labs can ship freely, but who they have to call first. The shift is small in any single decision and large in aggregate: a private lab that needs prior approval for new product is, by definition, no longer purely private.

Hand the agent a binary. Ask it to rebuild the program.

The team behind SWE-Bench released ProgramBench: 200 open-source projects where the agent receives only the binary and the documentation, and is graded on whether its reconstructed source passes 95% of the original test suite. Today's best score — Claude Opus 4.7 — sits at 3%. A useful new ceiling, and a humbling one.

Frame 01 · GPT-5.5-Cyber

A model only the defenders get to call.

OpenAI is widening its Trusted Access program for verified cybersecurity workers and shipping GPT-5.5-Cyber, a variant tuned for vulnerability research, exploit reasoning, and reverse-engineering work that the consumer model declines by default. The bargain: more capability, but only behind verified identity and audited use.