This week's usable signal is about prices and metrics rather than capability: what the automation costs, what the box costs, which number the comparison sorts on.
Nothing shipped this week that a non-developer picks up and uses on Friday. What did land is a set of numbers that change decisions already on the table: a promotional ceiling with an end date, a component price that repriced every self-hosting plan, and a measurement most model comparisons are still using wrong. Three items, all of them cheap to check.
Claude Code's elevated caps have an expiry (procurement).
Anthropic published the terms of its May to August 2026 weekly limits promotion, which HN read as notice that the elevated usage caps return to the standard ceiling. The distinction that matters for planning is between what a plan advertises and what a plan has quietly been allowing since May. Scheduled work is where this lands first: an agent pipeline running on cron does not complain when it hits a ceiling, it produces less and keeps its schedule. Anyone who sized a workflow against the promotional headroom is budgeting against a number with a date on it. The usage figures are already in the console. The ceiling they were measured against is the part that moves.
Memory prices just repriced the local box (procurement).
Tracked retail prices put 128GB of DDR5 at $3,399, roughly ten times its all-time low and up about 500 percent in a year. Last Thursday's edition covered the RTX Pro 6000 doubling to $16,000. Cheap system memory was the workaround for exactly that: buy ordinary RAM, push the model off the graphics card, accept slower tokens. The workaround now costs more than the thing it was working around cost a year ago. One counter-datapoint is real but narrow, a published configuration running DeepSeek V4 Flash at usable prompt speed on four used 12GB cards, about $1,000 of hardware. A small command-line utility, llmfit, built by AWS engineer Alex Jones, checks hundreds of models against a given machine's specs and reports which ones fit, which is a cheaper first step than a $3,399 order.
Tokens per second is the wrong metric (evaluation).
Testers running Qwen 3.8 27B report that it emits fewer tokens per second than faster-decoding alternatives and still finishes tasks sooner, with better results. Both things hold at once, because a model that reasons more per token spends fewer of them reaching the answer. Tokens per second is the number vendors publish and the number most comparison tables sort on, and for agentic work it measures the wrong thing. Wall-clock time to a correct result is the one that shows up in a bill and in a schedule. Changing that costs nothing: no different hardware, no different model, just the same comparison rerun with a stopwatch on the task instead of on the stream.
None of the three is a new capability. Each changes a number already sitting in a spreadsheet: the ceiling, the component cost, the column the comparison sorts on. The weekly usage figure in the Claude Code console is the cheapest of them to look at, and the only one with a date attached.