The AI Brief

CNBC reported July 7 that Chinese-origin AI models have held above 30% of US enterprise token volume on OpenRouter every week since February 8, peaking at 46%. The prior twelve-month average was 11%, and the figure was just 4.5% in H1 2025. OpenRouter itself expanded from roughly 5 trillion tokens per week in April 2025 to over 20 trillion by April 2026, meaning the Chinese share is growing on an expanding base. DeepSeek V4 Flash alone commands 16.3% of all OpenRouter token volume, more than any single US provider.

The price gap drives it. Open-source Chinese models are 60% to 90% cheaper than the leading Anthropic and OpenAI offerings, per Justin Summerville of OpenRouter. DeepSeek V4 Flash is priced at $0.14 per million input tokens; GPT-5.5 lists at $5.00. Z.ai's GLM-5.2, released by Chinese lab Zhipu AI in June under MIT license, saw the fastest adoption of any model tracked on Vercel in 2026: daily token volume grew approximately 27× and customer count 80× in its first full week. GLM-5.2 scored 62.1% on SWE-bench Pro, above GPT-5.5's 58.6%, while trailing Claude Opus 4.8 by one percentage point on leading agentic benchmarks. Coinbase runs roughly 1,200 AI agents and cut its AI spend approximately 50% by routing commodity work to GLM-5.2 and Kimi while reserving Opus 4.8 for hard problems. Airbnb and Uber have publicly acknowledged Chinese model usage; Uber burned through its full-year 2026 AI budget by April.

The revenue picture is less alarming than the volume picture: approximately 96% of teams running inference still keep an OpenAI or Anthropic account open, and Anthropic reportedly holds roughly 65% of Vercel spend even as raw volume drifts. The structural concern is forward-looking: the Fable 5 export-control episode in June accelerated Chinese model adoption directly, enterprises scrambled for alternatives during the 19-day outage with no transition period, and the underlying price gap predates that disruption and is widening.

Anthropic announced early on July 13, via its official @claudeai X account, that plan-included Fable 5 access on Pro, Max, Team, and premium Enterprise seats, and Claude Code's 50%-elevated weekly rate limits, are extended through July 19 at 11:59:59 PM PT. The announcement came after the July 12 deadline had already passed, making this the second extension following the first push (from July 7 to July 12, announced hours before that cutoff). No dedicated newsroom post accompanied either extension; both were made via X and updated support documentation.

The mechanics are unchanged: subscribers can apply up to 50% of their weekly usage limits to Fable 5 at no extra cost, drawing from the same shared pool as other Claude models. After July 19, Fable 5 exits plan limits entirely and moves to prepaid usage credits at $10 per million input tokens and $50 per million output tokens, the most expensive rate on Anthropic's current public price list, double Opus 4.8. Anthropic has stated the transition is a temporary capacity measure and that it intends to restore Fable 5 as a standard plan benefit when compute allows. An Anthropic deal with SpaceX using Colossus 1 data center capacity, adding over 300 MW and more than 220,000 NVIDIA GPUs, is cited internally as the mechanism for that expansion. Whether July 19 is a final deadline or a third extension point is not stated.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Thomas Bloom, a mathematician at the University of Manchester, provided the first substantive expert reaction to GPT-5.6 Sol Ultra's claimed proof of the Cycle Double Cover Conjecture, published by OpenAI as a PDF on July 10. Bloom called it "a very nice proof" that is "short, elementary, and could have been discovered in the 1980s." His read: the proof reduces the conjecture to cubic graphs, applies the 8-flow theorem, and uses linear algebra over GF(3), a finite field with three elements, to construct the required cycle cover. The core mathematical ideas trace back at least to a 1983 paper by Bermond, Jackson, and Jaeger. OpenAI's paper does not cite that work.

Bloom's critique is pointed but narrowly scoped: the strategy likely worked because the AI searched related literature and found the right combination of tools, then applied them without attribution, the same pattern that has recurred in other AI-generated proofs. He writes that "anyone reading only the paper might think the AI invented the underlying strategy itself." Bloom suspects the key step involved a small, counterintuitive twist in the linear algebra that human mathematicians would have abandoned after an initial failure; AI agents have no such discouragement heuristic and kept trying variations until one succeeded.

OpenAI's Codex engineering lead Thibault Sottiaux announced the result on X. The prompt OpenAI released directed the system to keep independent proof strategies alive simultaneously, route adversarial reviewers against surviving candidates, and reject any missing lemma as routine. Cost estimates for the run range from $275 to $485 at standard Sol API pricing. The Wikipedia entry for the conjecture was updated on July 10 to note OpenAI's claim. The Hacker News thread exceeded 200 comments within hours. No formal mathematical community verification has been published as of this edition.

Google DeepMind is targeting July 17 for the general availability of Gemini 3.5 Pro, per reporting from Business Insider and widespread industry coverage, first covered in Vol. I, No. 46. The July 17 date has not been officially confirmed in a Google newsroom post, model card, or API documentation page; as of July 13, the public Gemini API lists only gemini-3.5-flash and gemini-3.1-pro-preview. The model's internal development code involved scrapping the Gemini 2.5 Pro base and running a new pre-training cycle after enterprise testers flagged token-efficiency gaps, coding shortfalls, and reasoning failures the existing architecture could not close.

Reported, but unconfirmed, specifications include a 2-million-token context window (double the 2.5 Pro cap), a Deep Think Reasoning Layer for multi-step logic, and autonomous workflow capabilities for chaining tool-use tasks. Google is reportedly positioning the model as a cost-effective alternative to GPT-5.6 and Fable 5 rather than competing head-to-head on raw benchmark scores. Gemini 3.5 Flash, already shipping, scores 42% better than Flash 3 on Google's own long-range multi-turn cyber benchmark while achieving 68% better token efficiency. Unverified claims from private testing allege Gemini 3.5 Pro outperforms Fable 5; these should be treated as unconfirmed until independent evaluators publish results. The Gemini 3.5 Flash model is priced at $1.50 input / $9.00 output per million tokens; no Pro pricing has been announced.

Reuters reported July 7, citing three people familiar with the matter, that DeepSeek is developing its own AI chip focused on inference workloads. The project is in early stages: the company has contacted chip design, foundry, and memory firms, and has begun hiring semiconductor engineers. DeepSeek did not comment. The chip targets inference, the stage where a trained model serves responses to users, rather than training, where Nvidia's software stack and interconnect lead is widest and where US export controls bite hardest at the foundry level.

The strategic logic is clear. DeepSeek currently runs inference on Huawei Ascend accelerators after US controls blocked its Nvidia H800 supply. The company does not want to become dependent on Huawei either. A prior DeepSeek model release included a note that a training data format was "well suited for home-grown chips soon to be released", corroborating signals that the chip effort predates the Reuters report. DeepSeek launched a semiconductor design hiring drive in early 2025. The company is simultaneously raising its first-ever outside capital, reportedly $7 billion at a $52–59 billion valuation, with chip development expected to be a key use of proceeds.

Significant obstacles remain. Competitive AI inference chips typically take years and require advanced foundry access that US export controls restrict for Chinese designers. The report does not name a foundry, process node, or target first-silicon date, the unknowns that determine whether the effort is on a 2027 or 2030 horizon. DeepSeek joining OpenAI (Jalapeño, via Broadcom) and Google (custom TPUs) in custom silicon narrows the framing of the AI hardware market: the companies capturing the most value are increasingly those that own their inference stack end-to-end.