Issue 47: GPT-5.6 ships, TypeScript goes native, plus 3 picked for you
A frontier-model pileup, TypeScript's native compiler goes stable, and agents keep growing legs. Here is the week, ranked and cut to a 15 minute read.
Big moves
- GPT-5.6 clears its government review and ships, with a ChatGPT that does your work
OpenAI's limited preview is over. GPT-5.6 launches in three tiers, and ChatGPT Work blends ChatGPT and Codex into an agent that completes whole tasks across your apps and files. Altman calls the top tier the best model they have produced, and it is priced to undercut rivals.
The Verge · 4 min
- Apple sues OpenAI over stolen hardware secrets, and files the video to prove it
Apple's complaint alleges former employees who joined OpenAI's hardware unit took confidential files, and reportedly includes footage of one leaving headquarters. More than 400 ex-Apple staff now work at OpenAI as it races to ship its first consumer device.
CNBC · 5 min
- Grok 4.5 lands as an 'Opus-class' model at a fraction of the price
SpaceXAI's first model since going public is pitched as a cheaper, faster workhorse (about $2 in and $6 out per million tokens) that Musk says rivals Anthropic's Opus. It shipped the same week as GPT-5.6 and Meta's Muse Spark, in a suddenly crowded frontier.
TechCrunch · 3 min
The tools you use
- TypeScript 7.0 is stable, and the native compiler is about 10x faster
The Go-based port of TypeScript is production-ready, with roughly 8 to 12x faster full builds and a new LSP editor experience. One catch: Vue, Svelte, Astro, and Angular template checking has to stay on TS6 for now, so check your stack before you jump.
Microsoft · 4 min
- MCP's biggest revision goes stateless, and it is a breaking change
The new Model Context Protocol spec drops sticky sessions so servers can run behind a plain load balancer, formalizes an extensions framework, and graduates server-rendered UIs and long-running Tasks. There is a one-year grace period to migrate off the old surfaces.
Model Context Protocol · 5 min
AI and agents
- Google's Managed Agents can now run in the background and refresh their own credentials
Gemini's Managed Agents gained async long-running execution, direct connections to remote MCP servers, custom function calling alongside sandbox tools, and mid-session credential refresh, pushing them toward autonomous cloud workers that live inside a real dev environment.
Google · 4 min
- Meta opens its first paid model API with Muse Spark 1.1
Meta Superintelligence Labs released a multimodal reasoning model built for agentic tasks, with gains in tool and computer use and coding, and for the first time let developers build on it through a public Meta Model API.
Meta AI · 3 min
One deep read
- The data fingerprint of AI-era coding is now measurable
An analysis of 5,009 repos and 10,812 developers finds two independent signals of agent-assisted coding: explicit AI-tool mentions jumped from near zero to about 1 in 8 commits by mid-2026, and commit-message length shows a matching sudden, lasting rise. A data-first argument that this is a structural shift, not vibes.
jauu.net · 10 min
Quick hits
- OpenAI finishes the global rollout of GPT-Live, its more natural voice model, and doubles usage limits
The new voice generation can keep talking while it hands hard questions to a frontier model in the background.
OpenAI · 2 min
- Mercor buys agent-training startup Deeptune as its revenue crosses $2B and it eyes a ~$20B raise
Deeptune builds simulated environments to train agents; Mercor says annualized revenue passed $2B in June.
SiliconANGLE · 2 min
- Meta's own MTIA AI chips begin production in September to cut its Nvidia bill
Designed with Broadcom and built at TSMC, as Meta plans to deploy 7 gigawatts of compute this year.
TechCrunch · 2 min
- Prime Intellect raises $130M to let enterprises train their own agents
A full stack of compute, an RL framework, and evals so companies build agents instead of renting frontier labs. Customers include Ramp and Zapier.
TechCrunch · 2 min
- Ex-GitHub CEO Thomas Dohmke's new startup 'Entire' launches
A distributed network for mirroring code repos, pitched as infrastructure for a world flooded with AI coding agents hammering repos.
GeekWire · 2 min
Want an issue like this, tuned to you?
Start your free month