Meta shipped it September 2 on Muse Code and the Meta Model API with a 1-million-token context window and text, image, and video input, per its research blog. Pricing runs $1.25 per million input tokens, $4.25 per million output, and an 88% discount on cache hits; the model scores 48 on the Artificial Analysis Intelligence Index, 13th of 202 rated. Roughly half the input price of the reasoning models it benchmarks against is reason to run your eval this week, not next quarter.

MIIT released the plan on September 7 calling for 3.8 trillion yuan (US$532 billion) of cumulative information infrastructure investment through 2030, orderly deployment of clusters running 100,000 or more accelerator cards, and greater effort to adapt that infrastructure to home-grown chips, per the South China Morning Post. Capacity sat at 2,185 eflops at the end of June, so Beijing has committed to more than quadrupling it on domestic silicon. Price your 2027 accelerator supply against that number, not against today's export-control snapshot.

Yesterday's release lets background sub-agents inherit the parent agent's workspace sandbox, gives Instance AI search over its own past conversations plus folder browsing, and adds public API endpoints for Git push, pull, and source-control status, per the n8n release notes. Shared workspace state kills a lot of explicit node plumbing. It also means a sub-agent's blast radius is now the parent's entire workspace, so scope credentials at the sandbox, not at the node. Try n8n.

OpenAI shipped it September 8 with Sketch, which turns a rough drawing into a finished reference, plus better subject preservation from reference photos and steadier style across multi-turn edits, per OpenAI. Latency is down up to 50% against Images 2.0, and two API models landed with it, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. If your pipeline burns a human minute waiting on each revision, that halved wait is the whole upgrade.

GitHub's 67% cost-reduction headline holds on exactly one of the three benchmarks it published; a second cuts cost 65% for a rounding-error quality change, and the third saves 36% while giving back 1.5 quality points, per GitHub's writeup. That spread is the entire decision, and our guide turns it into a two-row rule plus a modeled monthly bill. See which bucket your work falls in. Get Copilot.

None of the three opt-in controls change model behavior, API pricing, or rate limits; what changes is who holds the logs and whose security team gets the flags. Our writeup walks each control and the procurement objection it actually answers. Read the control-by-control take.