Three prices, one model layer. DeepSeek answered $65B with open weights. Apple slipped Siri to 2027. Capability is now ambient.
‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌
to make Anthropic the most expensive frontier lab in history. DeepSeek shipped a peer-tier flagship model for
and made open weights the cheapest. Apple bet on Gemini — and pushed its launch to
The closed-model thesis assumed capability was the moat. This week proved it isn't. Hyperscalers paid for exclusivity; DeepSeek made it worth less; Apple paid for Gemini and discovered its chip couldn't run it. Three companies, three different problems — none of them solved by spending more on the model.
on Saturday — a 1.6-trillion-parameter mixture-of-experts model with MIT-licensed open weights, trained on Huawei Ascend 910C clusters for
Hugging Face downloads in the first 12 hours, less than a day after Google's
Anthropic commitment closed. The closed-model alliance was structured to make price competition unwinnable; the open-weight response got faster overnight.
The closed-model deal took five days and $65 billion. The open-weights answer took twenty-four hours and $14.7 million. The race for the model is no longer where the moat lives.
told analysts on Sunday that the Gemini-powered Siri rebuild won't ship until iOS 27 in fall
— a year later than originally targeted — with two suppliers telling
the actual cause is a chip-level redesign of the Neural Engine in next year's A20, not the privacy-controls explanation Apple gave publicly.
Apple chose Gemini to power Siri — then realized its chip couldn't run it. The 2027 date isn't a privacy delay; it's a Neural Engine redesign.
at Dev Day Saturday in Memphis — a 600-billion-parameter dense model with a
context window running on Colossus 2 — priced at
per million output, undercutting GPT-5.5 by
Grok's 280 million DAU isn't capability — it's distribution. Being one tap inside X is now a cheaper growth lever than another billion in compute.
's Pentagon contract surfaced Friday in House Armed Services records — a
two-year award for the Joint All-Domain Command and Control program covering GPT-5.5-Defense across Combatant Command planning, with explicit carve-outs around autonomous weapons targeting. Three OpenAI policy researchers resigned within hours of disclosure.
The $480 million is the headline. The signal is that three OpenAI policy researchers resigned within hours. OpenAI is becoming a defense vendor — and the resignations are saying it before any press release will.