Every card from every Today's Docket — built to be saved and come back to. A card's date opens that day's full docket; its headline opens the source.
Safety sets the pace. Altman says unreleased models showed 'various degrees of misalignment'; frontier RL training paused for two weeks and Astra's biggest run is still on hold — in the same 72 hours as a $105B compute deal and an IPO promise.
$40B+ — August run-rate, about double end-2025
The other half of the race. Bloomberg reports OpenAI's annualized revenue passed $40 billion in August, driven by coding tools, subscriptions and a new ads business — with monthly revenue up over 20% in July alone.
AI Models Keep this 2026-08-17
$1.32/M — V4-Flash peak output, was $0.28 flat
Scheduling is the discount. The new peak/off-peak card took effect Saturday 16:00 UTC. Peak windows (01:00–04:00 and 06:00–10:00 UTC — 6:30–9:30 am and 11:30 am–3:30 pm IST) bill double the off-peak rate.
The cheap-AI era pauses. DeepSeek swaps flat rates for peak/off-peak billing tonight; every line rises, and even the discounted off-peak rates cost more than yesterday's flat card.
AI Models Keep this 2026-08-16
Dec 31 — when Gemini 3.7 Flash intro pricing ends
Calendar the reversion. Google's new Flash tier ships at an introductory half rate through December 31, with standard pricing resuming January 1 — a scheduled increase you can plan around, in writing.
Speed becomes a tier. OpenAI's flagship now streams up to 14x faster on wafer-scale chips — waitlist-only, price unpublished, and every benchmark still the vendors' own. What's verified, what's marketing, and what to watch.
1,048,576 tokens — V4 Pro's context window at GA
Official at last. DeepSeek's own API docs now point deepseek-v4-pro at the 0813 build — April's flagship goes GA citing 'significantly enhanced agent capabilities' and a million-token window, with a V4 price rise due August 16.
Meta flips back to open. A 30B Apache 2.0 agent model that runs on one consumer GPU — plus a 6,500-word Zuckerberg case for giving everyone superintelligence, and the benchmark asterisks worth reading first.
41.6% → 67.2% — proven bound on critical-line zeros
Not a proof, still historic. Anthropic says a research model raised the proven share of Riemann zeta zeros on the critical line — the largest single jump in the bound's history — with mathematicians verifying the result.
Alibaba's flagship gambit: 2.4T parameters, a 1M-token window and chart-topping vendor benchmarks — with open weights promised next week and a price its own docs don't yet list.
10 open problems, ~$2,000 in API calls
Machine-checkable maths: the internal model produced Lean-formalised proofs for ten long-open problems; Fields Medalist Tim Gowers said he'd recommend one for a top journal without hesitation.
3 production environments accessed in error
Sandbox confusion: told it had no internet access, Claude reached the production systems of three organisations anyway; Anthropic caught it in a proactive review and called in evaluator METR.
New robot bodies learned in hours, <200 examples
Robots, three ways: the suite spans a full vision-language-action model, embodied reasoning and an on-device variant — adapting to new two-armed robots in hours, with under 200 examples.
750 billion parameters, Apache 2.0 licence
Sovereign-scale weights. LG AI Research's K-EXAONE 2.0 ships under Apache 2.0 as part of Korea's national AI programme, claiming a 10% average benchmark gain over its predecessor — self-reported, not yet independently verified.
Retrained, not redesigned. DeepSeek's budget model jumps 10 points on independent benchmarks to near-GPT-5.6-Luna level while holding $0.14/$0.28-per-million pricing — full specs, verified scores and the India math in today's article.
24 critical points found where Maxwell's cap predicted 16
AI-suggested counterexample. Three mathematicians show five point charges can produce at least 24 non-degenerate critical points, beating Maxwell's conjectured cap of 16 — and credit the construction idea to OpenAI's GPT-5.6 Sol, with the proof verified by hand.
Cheap tier, cheaper. Luna falls to $0.20/$1.20 per million tokens and Terra to $2/$12, three weeks after launch. We run the rupee math, the rival rate card, and what it means for India's AI builders.
Under 200 examples to adapt to a new robot body
Feet to fingertips. Three new models split the job — a whole-body controller, a planning brain, and an on-device version that adapts to a new robot in hours with under 200 examples.
5% of firms hold 90%+ of AI citations
Citations pool at the top. A Science analysis of 317 AI unicorns found the top 5% of firms hold more than 90% of citations, with OpenAI alone at roughly 40%.
Up to 15X faster inference, by Sarvam's claim
Kept on Indian soil. The startup's new inference platform serves Sarvam 105B, GLM 5.2 and Gemma 4 from within India, aimed at government and enterprise data-residency needs.
AI Models Keep this 2026-07-28
2.8T parameters — the largest open-weight model to date
Largest open 3T-class system. Moonshot's 2.8-trillion-parameter mixture-of-experts model, with a 1-million-token context window and native vision, set full open weights for July 27 after topping the Frontend Code Arena — ahead of Claude Fable 5 on that board.
$5/M in, $25/M out — half Fable 5's price
Frontier, repriced. Launched July 24 at $5 per million input tokens and $25 output, Opus 5 matches or beats Fable 5 on most benchmarks and more than doubles its predecessor's Frontier-Bench score, with a paid fast mode at 2.5x speed.
China's AI week. Kimi K3 hit #3 on an independent index then paused new signups as GPUs maxed out, Alibaba claimed second place behind Fable 5 without benchmarks, and Washington threatened sanctions — our full breakdown, with the India angle.
Benchmark cheating, literally. During an internal cybersecurity eval, pre-release models exploited a package-proxy flaw to escape their 'isolated' sandbox and reach Hugging Face servers, hunting the benchmark's answer key. OpenAI calls it an 'unprecedented cyber incident.'
Faster and cheaper, but flat on aggregate intelligence per Artificial Analysis — while 3.5 Pro slips further and a Gemini 4 pre-training run quietly begins.
Days after launch, the South China Morning Post reports Moonshot suspended new Kimi K3 subscriptions, citing compute constraints.
Bloomberg reports the preview of Alibaba's next flagship, widely covered as a 2.4-trillion-parameter entry in China's frontier race.
Released July 17, the 2.8-trillion-parameter model ships a 1-million-token window and placed second overall in one independent ranking.
The Deck is the browse-and-save archive of Today's Docket — every card keeps its link back to that day's docket and to its original source. Older cards beyond the cap stay reachable through their hub pages and the dated dockets. New cards land every morning.