Cheaper tokens and tighter seat limits landed the same day. The rest of this week's usable items are questions to put to a vendor, not tools to install.
DevDay produced the week's only unambiguous operator news, and it arrived in two halves that most coverage treated separately: cheaper tokens, tighter seats. The rest of the week ran developer-deep, mostly local inference configs and memory-bandwidth numbers. Four items below survive the translation to operator altitude, and three of them are questions a vendor either has an answer to or does not.
OpenAI cuts token prices, tightens seat limits (procurement).
GPT-6.1 Sol shipped at DevDay priced around 20% of Astra, with quality claims that are OpenAI's own and verified by nobody outside the company yet. The same day, the $200 ChatGPT Pro tier's limits were halved, and a new $500 tier now sits roughly where the old $200 allowance did. Per-token pricing improved; per-seat pricing got worse. For teams that deliberately pushed recurring work into human-operated chat seats to avoid API bills, that calculation has flipped. The arithmetic is visible without touching code: seat count times price, against last month's actual usage per person, against what the same volume costs per token.
Always-on agents become a purchasable shape.
Dots are agents that stay running instead of waking up per request. Most automation sold as an AI assistant today works the other way: a request arrives, the model answers, the state disappears. Continuous operation is a different product shape, and it carries a failure mode already visible in other always-on deployments, where an agent replies on your behalf from a picture of the world that went stale hours ago. The question worth putting to a vendor is whether their automation runs continuously or per request, and what it does when its state is out of date. Buyers who have been paying for agents that are really scheduled jobs will find the distinction clarifying.
A distilled model ships a disagreement bound (evaluation).
A single developer released a tool that trains a small local model to imitate Jev's decision-making, and shipped something unusual alongside it: a stated bound on how often the small model decides differently than the model it copied. That is a ceiling on divergence rather than an average score, and it is a procurement vocabulary upgrade available today whether or not the tool itself is relevant. Any vendor selling a cheaper, smaller, faster version of a model you already trust can be asked the same thing: how often does the cheap one disagree with the expensive one, and what is the number. Benchmark averages hide exactly that. Most vendors will not have it, which is itself informative.
Anthropic publishes numbers on a downloadable model.
Anthropic's Frontier Red Team measured GLM-5.3, an open-weight model anyone can download, producing full control-flow hijacks (working exploits, in plain terms) on 4% of a random 100-task sample of binary exploitation problems. Anthropic's own report calls it a threshold crossing: the two prior models it ran through the same tasks, Claude Opus 4.6 and GLM-5.2, scored zero. Simon Willison's writeup is the readable curation of that finding, not an independent take. The percentage matters less than the fact that a frontier lab now publishes capability measurements on a competitor's downloadable weights, in the same week that talk of restricting Chinese open weights picked up again. Those two facts sit together for anyone buying AI tooling: measured capability rising is also availability risk rising. If a vendor in your stack is running GLM underneath, the answer to what happens when that model becomes unavailable is better to have before the question is urgent.
Three of these turn into a vendor conversation. One turns into a spreadsheet, and it is the one with a date on it: the new ChatGPT Pro caps hit the next billing cycle, and the single number that settles whether seats or tokens win is last month's real usage per person. Most teams have never pulled it.