The lever in both arguments is the same lever. The 75.5-point accuracy gain on the CAD reconstruction task came from adjusting weights the publisher no longer controls [11][4], and adjusting weights is the first route the research report quoted by Nimit Mehra names around a model's built-in safeguards [3]. A buyer who wants the first performs the second as a matter of arithmetic, not intent.
The mechanical difference between the two deployment routes is narrower than the debate and more consequential than it sounds. Under API access, customization happens only through the capabilities the provider chooses to expose, and the weights stay on the provider's side [5]. Under an open-weight license, the artifact runs on infrastructure the customer controls and can be adapted with the customer's own data [4]. Everything about who tests what, and who can withdraw what, follows from that one distinction.
The cost case is real and reported narrowly. EnterpriseLab's specialized 8-billion-parameter model matched GPT-4o while cutting estimated inference cost by a factor of eight to ten [9], which is 87.5% to 90% off that one line [14]. Estimated inference cost is the only figure the source gives; the training run, the evaluation harness and the re-testing each subsequent tune requires sit outside it. We do not know from this record what the tuning bill was.
A skeptic would note that these numbers arrive in a contributed column by a founder who sells AI workflow software [10], and that CAD program reconstruction from images is one narrow task. Both true, and they cap how much weight the 9.9-point margin over GPT-5.2 should carry [12]. What does not depend on any benchmark is the definitional point about control of the weights [5], which is why the procurement question survives even if the accuracy figures do not replicate.
Then there is the reason anyone is tuning at all. On the FinBalance benchmark, which asked models to create journal entries, reconcile them and produce a balance sheet, the best of six LLMs reached 46% [6], leaving 54% of a multi-step workflow unaccounted for [13] in a domain where reasoning models have been passing the CPA exam since 2023 [7]. Credentialed knowledge is not the same as reliable end-to-end execution, and the source is explicit that the models produced plausible entries while failing to carry them across supporting documents [6].
Sequencing is where this quarter meets next. A decision to close the domain gap by tuning open weights sets up a standing obligation to re-run safety evaluation after every tune, because the tested artifact and the deployed artifact are no longer the same object [3]. That obligation is cheap to accept in a pilot and expensive to discover in year two, and the record here does not price it.
