The million-token model nobody will claim is keeping what you send it

The capacity number is the part worth arithmetic. OpenCode put the provider's headroom at 100 trillion tokens a day ^[3]. That is about 1.16 billion tokens every second ^[10], or roughly 100 million complete fills of the advertised million-token window in a single day ^[11]. Someone is covering the inference bill for that and charging nothing for a week ^[3].

The listing's wording repays a slow read. The assurance is that retained prompts and completions are not used for training ^[2]. Training is the whole of the promise. Retention period, access control and storage location are not addressed, and there is no counterparty to ask, because the provider has not said who it is ^[1].

That gap is where the compliance problem sits rather than in model quality. European data protection law wants a contract with a named processor and an assessment of where the data ends up, and an anonymous supplier can satisfy neither ^[7]. The attribution guesses do not help: the leading theory names Z.ai, which has form here after testing GLM-5 anonymously under another name, while a tokenizer analysis points instead at Microsoft's MAI family ^[5]. Those two answers imply different places for your source code to land. The AI analyst Andrew Curran, writing over the weekend, said people seemed "less sure of anything" than they had been the night before ^[6].

The regulatory clock is already running. The AI Act's transparency obligations took effect on 2 August, with penalties reaching 15 million euros or 3 percent of global turnover ^[8]. The 3 percent limb only overtakes the fixed figure above about 500 million euros of turnover ^[12], so for most firms using this thing the binding exposure is the flat cap, and the regime it belongs to is built on knowing which provider is responsible for what ^[8].

Meanwhile the adoption is real and the reviews are good. Stripe's chief executive Patrick Collison called Ox Alpha "very impressive", and it is pitched at coding, long-horizon agent work and production use ^[4]. The model is being tried across the industry despite nobody taking credit for it ^[14]. TNW's own read is that a stealth launch is a legitimate way to benchmark before announcing, since open-weight releases have closed the capability gap faster than the safety one ^[13], and that the sensible European position is to test it with nothing that matters ^[9].

Strip out the mystery and the transaction is legible enough. An unnamed party is buying a very large sample of real engineering prompts, at a price it has chosen to pay in compute, and the only published term of the deal is that it keeps what arrives ^[2][3]. Free capacity of that size has to be worth something to whoever provisioned it, and the retention line is the only clue on offer as to what.