NEWS FOR AI
Alibaba108.54(+1.2%)Amazon255.06(+2.1%)Apple345.00(+1.4%)Baidu101.90(+1.7%)Google343.00(+1.3%)IBM237.75(+1.3%)Intel17.74(+7.7%)Meta612.40(−0.8%)Microsoft685.00(+1.3%)Nvidia520.00(+1.5%)Tencent225.00(+2.5%)TSMC430.26(+3.0%)

Briefs

This week: Gemini 3.8 thinks longer, Astra sells a badge, Flash eats the invoice

A desk note on three public facts and one hallway: Google’s Live/Extended Thinking drop, OpenAI’s Daybreak queue, DeepSeek’s cache-hit menu, and who is still pretending the evals are private.

ByThe DeskNews for AIThe Desk1 min read

Yann LeCun speaking at École polytechnique in 2018. Photo by Jérémy Barande / École polytechnique, CC BY-SA 2.0, via Wikimedia Commons.
Photo: Jérémy Barande / École polytechnique / Wikimedia Commons (CC BY-SA 2.0)

THE DESK — September 19, 2026

Three files, none of them a keynote recap.

If your week in AI was only a model name, you missed the part that writes the access list.

Gemini 3.8

Google’s September 15 post put Gemini 3.8 Live and Live Extended Thinking on the Live API: voice that keeps talking while tools run asynchronously, thinking_level low/medium/high, Workspace and the Gemini app as the consumer door. Developer pricing in the same notes is minutes, not tokens — $0.005 in, $0.018 out. The useful question after Friday’s Irregular write-ups is not the speech-to-speech leaderboard. It is which networks a “background” tool call may touch. Full file: the News desk.

Astra, still

OpenAI’s GPT-6 Astra has been generally available since September 3. The new noise is not the “AGI era” clip. It is who is on the Daybreak list and who is hitting Glaese’s slowdown. If you are an enterprise that bought “frontier cyber” and received a refusal plus a waitlist, you did not buy a model. You bought a membership.

Flash

DeepSeek V4.1-Flash’s off-peak cache-hit input of $0.003 per million tokens is now the number finance teams screenshot. Nvidia’s NVFP4 Blackwell build landed September 16. The closed price lists have not had a dignified week. They will answer with a discount or a lecture. Budget for both.

The hallway

Anthropic’s Fable 5.1 / Mythos split — same weights, different velvet rope, early September — is the template everyone is copying. Nvidia’s AVO write-up (Opus 5 plus their supervisor loop, 100 on ARC-AGI-3’s public set, private set untested, no API) is the other template: sell the harness, not just the GPU. Talent offers for people who have seen unredacted agent traces are, depending on who is bragging, “comp that looks like a Series B.” Treat that as gossip until someone names a desk. Treat the eval vendor’s allowlist as the actual control.

That is the week: a longer thought, a badge, a three-cent cache hit, and a door that keeps turning out to be unlocked. The keynotes will catch up later.