Two things happened to the AI frontier this week, and they point in opposite directions.

The first is speed. Inside a single 24-hour window, three labs pushed the capability-per-dollar line further than it has ever been. The second is doubt: a letter now carrying 1,376 signatures from inside those same labs asks the US government to help build the tools to deliberately slow the whole thing down. Both are real, both landed in the same fortnight, and the gap between them is the story.

The race got cheaper, not just faster

On 12 August, SpaceXAI — the company formerly known as xAI, since folded into Elon Musk's SpaceX — released Grok 4.6. On the independent Artificial Analysis Intelligence Index it scores 61, level with OpenAI's GPT-5.6 Sol and a step behind only Anthropic's Opus 5 (63) and Fable 5 (62). That is roughly the top of the public leaderboard. The price is the part that matters: $2 per million input tokens and $6 per million output, unchanged from Grok 4.5, and well under what the models it now matches charge. SpaceXAI has pointed the release squarely at long-running agents and interactive, visual coding work — the workloads that burn the most tokens, and where a halved bill compounds fastest. Musk is already teasing a 4.7 "weeks away."

The same day, DeepSeek moved its flagship V4 Pro out of preview — a 1.6-trillion-parameter mixture-of-experts model (49B active per token, a one-million-token context window) shipped as open weights under the permissive MIT licence. The Pro repository on Hugging Face has logged more than 1.4 million downloads in the past month. Its listed API price sits several multiples below the US frontier — by DeepSeek's own accounting, tens of times cheaper per output token than the flagships it benchmarks against. Two frontier-tier drops on one day, both undercutting the incumbents on price: the word doing the rounds is "price war," and for once it fits.

The third name is the cautionary one. Alibaba's Qwen3.8-Max — a 2.4-trillion-parameter multimodal model — went live via API on 3 August and was billed as Alibaba's first open-weights release of a Max-class model, with the weights promised for "the week of 10 August." That week has passed. As of today the Hugging Face page is empty and no licence has been named. It is a useful reminder that "open" is, for now, still partly a press-release word. DeepSeek shipped the weights; Qwen announced them.

A billion people, and no subscriber number

Adoption kept pace with the capability. On 11 August, Sundar Pichai said the Gemini app had crossed one billion monthly active users — the fastest any Google product has reached that mark, up from 400 million in May 2025. It puts Gemini alongside ChatGPT, which passed a billion in June, and turns the consumer assistant race into a two-horse contest at real scale. Worth noting what Google did not say: how many of that billion pay. A monthly-active number is a reach figure, not a revenue one, and the omission is the tell.

And, quietly, the brakes

Against all of that sits "Pacing the Frontier," an open letter that has been gathering signatures since late July and now stands at 1,376 engineers and researchers from the labs building these systems — Anthropic, OpenAI, Google DeepMind and Meta among them. The signatories are not fringe: Anthropic's Dario Amodei, Jared Kaplan and Jack Clark; OpenAI chief scientist Jakub Pachocki; Meta chief scientist Shengjia Zhao; Google's head of AI safety, Anca Dragan. Their ask is narrow and specific — that Washington support an international effort to build the technical and governance tools needed to pace automated AI development, so that a coordinated slowdown becomes possible at all. Their fear is narrower still: that the labs are close to automating AI research itself, the point where models start improving models faster than humans can follow, and that competitive pressure makes any single company stepping back close to impossible.

The read

Put the two halves together and the tension is obvious. The people asking for a brake are, in many cases, the people shipping the accelerator. That is not hypocrisy so much as the exact problem the letter describes: no lab can slow down alone without handing the lead to whoever does not.

And notice which part of this week a brake could never reach. A closed model can be rate-limited, priced up, or held back. Open weights cannot be recalled — once DeepSeek V4 Pro is on a million machines, no letter and no ministry un-releases it. The cheapest, most copyable frontier is also the least governable one, and it is the axis moving fastest. The signatures are worth taking seriously. So is the fact that the week they climbed past 1,376, the frontier got measurably cheaper to run.