Four defaults worth rechecking this week: what a hosted agent retains, how long a latency number stays true, what a batch workload costs in hardware, and what a shipping model trades away for throughput.
Four items this week each undercut a number or a default an operator is probably still pricing against. None of them require touching a serving stack to matter. Three change what a reasonable question sounds like in a vendor conversation, and one changes what a batch workload costs.
A priority dispute over unpublished drafts (policy).
Tristan Buckmaster, an NYU mathematician, and Levent Alpöge, a researcher at Anthropic working in a personal capacity, say they fed a year of unpublished drafts into Codex, OpenAI's hosted coding agent, while developing their own solution, and that OpenAI announced a result on the same problem days before their planned publication. Buckmaster has publicly questioned whether OpenAI's model benefited from that private Codex data, a question OpenAI has not directly answered. The facts are contested, unresolved, and may stay that way. The operational question does not wait on the resolution: whether the default contract on an AI account retains prompt context, and for how long. Zero-retention endpoints, where the provider commits to not storing prompt or output data past the request, exist at most major vendors, but they are typically opt-in and often tied to a specific tier or a signed addendum. That distinction is answerable in one email to whoever holds the contract. Most accounts sit on the default.
Latency numbers have a shelf life (procurement).
A hands-on comparison ran the same model, Qwen3.8-Flash-Next, through three serving engines at full context and measured time-to-first-token at 35 seconds on the fastest and 258 seconds on the slowest. Same weights, same prompt, 7x apart. The spread traced to code changes still in flight, not to configuration anyone chose. Last Thursday's edition, The Layer Around the Model, covered the variance the serving layer introduces. The new part is how fast it decays. A time-to-first-token figure in a vendor deck describes software that has since been patched, and an internal benchmark from a month ago describes a build that no longer ships. When a quote cites latency, the follow-up that gets somewhere is when the number was measured, and against which engine.
A 2.8T model off four consumer SSDs (cost).
deltafin runs Kimi K3, a 2.8-trillion-parameter model, on a single MacBook Pro by streaming pieces of it off four consumer SSDs instead of holding the whole model in memory. It produces roughly one token per second. The architecture is mixture-of-experts: only a small fraction of the parameters activate for any given token, so the constraint moves from memory capacity to how fast the drives can feed it. That is about $400 of storage standing in for tens of thousands of dollars of accelerators. Nothing a person waits on survives one token per second. Offline work does: evaluation runs, dataset labeling, overnight synthesis passes where output quality matters and the clock does not. If a batch workload was priced against GPU capacity, the floor under that price moved this week.
Diffusion decoding lands in a shipping model (evaluation).
Inception Labs shipped Mercury 2.5, a language model that generates text in parallel passes rather than one token after another, left to right. Nearly everything in production today works the second way, and that sequential shape is where a good deal of the latency lives. The trade is quality headroom for throughput, which makes Mercury the wrong pick for work where a frontier model earns its price and a plausible one for high-volume completion work that is currently overpaying a frontier model to do something routine. Classification, extraction, autocomplete, form-filling. The evaluation is the ordinary one: same inputs, both models, compare the outputs and the cost per thousand calls. The architecture is unusual. The procurement question is not.
Of the four, one has a same-day answer. Whoever holds the AI contract either knows whether the account sits on a zero-retention endpoint or can find out before lunch, and the answer determines what work is safe to paste into a hosted agent between now and the next renewal. The other three need a benchmark, a weekend, and a real workload to point them at.