The week presented itself as a model week. Google shipped Gemini 3.8 Flash alongside a security-tuned Cyber sibling, DeepSeek's experimental vision variant went from announcement to runnable local weights inside two days, and Meta's Muse Spark 1.3 arrived with open weights signalled as close behind. Read only the release notes and the industry looks like it is still competing on weights.
Then a benchmark called FrontierHarness Eval ran identical models through nine different agent harnesses and reported cost per successfully completed task. The claimed spread is roughly 17x. A harness is the scaffolding that owns the tool loop, the context management, and the retry policy: for most teams it is code someone wrote in an afternoon many months ago and nobody has opened since. If the multiple holds even approximately, it dwarfs every model-swap decision currently sitting on a roadmap.
The model is one component in an assembly now, and the assembly is where the variance lives. It is rarely the most expensive component, and it is almost never the one with a launch post attached. Once you are looking at the assembly rather than the weights, most of the week's signal rearranges itself.
The harness mechanism is legible once stated: scaffolding that burns three failed attempts at half the token price costs more than scaffolding that lands on the first attempt, and every cost-per-token dashboard in production will report the first one as cheaper. The methodology deserves scrutiny before anyone reorders a vendor list on the strength of it, since harness benchmarks bend hard to task mix. The measurement axis survives that caveat intact. Cost per passing task is a number most teams do not currently produce, and the operator question is who in the building would even own it.
Simon Willison's datasette-mcp 0.2 changed the shape of SQL results from arrays of arrays to arrays of objects, explicitly because weaker models mis-align columns against positional rows. That release note is a design lesson about the same layer: the shape of what a tool returns is prompt engineering, and self-describing output beats token-efficient output whenever the model on the other end is not the strongest one available. The model did not change; the task success rate did.
The local-serving thread runs in the same direction. Someone grafted a Q8 n-gram layer into a Q4 Qwen quant and measured quality up with speed flat. New AP quantization formats surfaced for Qwen3.8 Flash with no benchmarking attached to them yet. Perplexity published the Mac inference server it runs internally for Qwen 3.6, which is production scheduling and batching code rather than a weekend wrapper. The artifact you actually run is assembled: components at different precisions, served by code somebody maintains, wrapped in a harness somebody wrote.
Jensen Huang put the vendor side of this plainly at CES: "You sell a chip one time, but when you build software, you maintain it forever." The buyer side is the mirror image. The model you rent is a line item you can change in an afternoon; the scaffolding around it is a maintenance obligation that accumulates quietly and never appears on an invoice. I have watched this shape arrive before. By the mid-2000s the storage vendor on the purchase order mattered less to most shops than whoever had tuned the queries hitting the array, and the purchase order was still where all the arguments happened.
The grounding layer went adversarial in public. An investigation published at Trellner found three operators mass-producing 215,128 fabricated "best software" pages built specifically to be retrieved and cited by AI search, with Perplexity citing them. Monday's edition asked who gets to say what shipped; this is that question one layer down, where the answer turns out to be that anyone with a template farm and an economic motive gets a vote. Retrieval is an input the model does not control and mostly cannot audit.
Two numbers carry the week, and neither one is a benchmark score. 17x is the distance between the best and worst way to run the same weights. 215,128 is the count of pages built to be read by a retrieval layer most teams inherited rather than chose.