The Omniscient Bulletin · 2026-08-13
The Omniscient Bulletin — August 13, 2026
Three frontier models shipped in one afternoon, DeepSeek's arriving with no announcement at all at a fraction of Grok's price; Anthropic is talking about a $6 billion acquisition; and Reuters says the next flagship Gemini is two months late.
Good Morning! Wednesday compressed a quarter's worth of frontier releases into one afternoon: xAI staged a launch for Grok 4.6, DeepSeek shipped its flagship silently through an API docs line, and Alibaba put Max class weights in the open for the first time. The through line is that the scoreboard is now the product: price per token, benchmark tables, and who gets to publish them. Washington noticed the open half of that race the same day, and Google, which put Gemini on a billion users a day earlier, turns out to be running two months behind on the next one.
Frontier
Three frontier models shipped in one afternoon, and the quietest release undercut both loud ones on price
xAI announced Grok 4.6: 61 on the Artificial Analysis index, matching GPT-5.6 Sol Max and one point behind Fable 5 Max, at $2 in and $6 out per million tokens, the same rate Alibaba set hours later. Within hours, DeepSeek shipped the GA build of its V4 Pro flagship to its API and OpenRouter with no blog post or tweet, just an API docs line, at $0.435 and $0.87 for a million tokens of context. By evening Alibaba had put the 2.4 trillion parameter model behind Qwen3.8-Max on Hugging Face, the first time its Max tier has gone open weight.
Frontier
Google formalized its DeepMind succession, and the next flagship Gemini is two months late
Reuters got inside Google's AI reshuffle: the next flagship Gemini has been delayed by two months after internal testing kept lagging rivals on coding, and Sergey Brin spent months pushing DeepMind to move faster, urging hundreds of employees at an April town hall and steering resources toward recursive self improvement. Koray Kavukcuoglu, chief AI architect since 2025, now has final say over DeepMind and keeps reporting to Pichai; Hassabis becomes chair with fewer management duties. Alphabet fell 4 percent the day the change was announced.
Industry
Anthropic's largest ever acquisition talk is a $6 billion bet on making its GPUs go further
Bloomberg reports Anthropic is negotiating to buy Decart for about $6 billion; the talks are not final and could still collapse. Decart, founded in 2023 by Dean and Orian Leitersdorf and Moshe Shalev, does real time generative video and, more to the point, software that makes GPUs train and serve models more efficiently, which Bloomberg frames as letting Anthropic's existing infrastructure absorb more demand. Decart raised at almost $4 billion in May, so the price is a roughly 50 percent markup, ahead of Anthropic's anticipated IPO.
Industry
Apple wants to pay publishers per use, not per year, to feed the new Siri
The Journal reports Apple is negotiating multiyear deals with news publishers to give the rebuilt Siri access to current events, with a twist on the standard template: variable compensation, paid when a publisher's content is actually used in a response, with a possible nine figure budget discussed. AI content deals to date have mostly been flat fees for broad access. For publishers already split between suing AI companies and licensing to them, usage metered pay is a third door.
Policy
The frontier framework the public cannot read is about to cover open models too
Wired reports the White House is 'almost certain' to expand its AI framework, the unpublished arrangement under which the most powerful US models get federal safety testing before release, to open weight models once they reach the frontier capabilities of Anthropic's Mythos class or GPT-5.6. Officials worry a closed only regime would brand open models second tier and scare enterprises off them, and equally that a 30 day testing gate could stifle what it covers. The framework stays voluntary for now.
Labor
Twitch turned on AI training for every streamer, and its product chief said the quiet part
Twitch added a 'training for generative AI' toggle to channel settings, default on, covering stream recordings for training Amazon's generative models; creators who object have to find the switch under security and privacy and turn it off. Asked on an official stream with about 3,000 watching why it was not opt in, chief product officer Mike Minton answered: 'If this was opt-in, nobody would opt in. That's honestly the answer.' The default plus the admission is the story: training data consent is now being designed around the expectation of refusal.
Compute
Nvidia's own store now wants $16,000 for a card that took pre-orders below $8,000
Nvidia's US Marketplace silently repriced the 96GB RTX PRO 6000 Blackwell at $16,000, the second hike this year: the card launched at $8,565, took pre-orders as low as $7,673 in March 2025, and moved to $13,250 earlier in 2026. Retailers have not caught up; Newegg lists the same card at $13,998 and B&H at $15,499. Workstation cards barely have a real MSRP, which is what makes the list price a useful signal: it is the shortage, memory prices above all, showing up in Nvidia's own posted number rather than in scalper listings.