Cheaper,
not catching
up.

Backlog #97: “The open-vs-closed model price gap over time (extends run 001) — [TREND] · `compiled`.” Run 001 fit the floor and the frontier of LLM API pricing over time, pooled across all models; it didn't split by open-weight vs closed. This run does, using a different, keyless source: OpenRouter's own public model catalog, GET /api/v1/models — no API key, no account, a plain GET request that returned 460 models today, before any filtering. Open-weight is OpenRouter's own binary flag: a published Hugging Face id (152 of 358 models in the final sample). Price is a blended $/million-token figure, the standard 3:1 prompt:completion weighting price trackers use when a single number is needed. “Release date” is this API's own created field — when OpenRouter listed the model, not independently verified against each lab's own announcement date, stated as a limit.

Three filters, all named rather than silently applied: 90 of 460 catalog ids carry a :free or :batch suffix — a second commercial tier of a model already listed under its base id, not a separate release, dropped to avoid double counting (8 models exist only under a suffixed id and are lost entirely, not recovered under another name). 4 rows price at exactly zero on both prompt and completion — not a real market price, dropped. 8 rows — all meta-routers that pick a model per request (openrouter/auto and its variants, nvidia/switchyard, typesafe/jev-router) — carry this API's own -1/token sentinel for “variable, no fixed quote,” not a real negative price; dropped. 358 models remain: 152 open-weight, 206 closed.

The gap itself is large and real. Median blended price today: $1.975/Mtok closed vs $0.352/Mtok open — closed models price a median 5.60× above open-weight ones, bootstrapped (4,000 draws) 95% CI on the log gap [+0.5740, +0.9566], or [3.75×, 9.05×] — excludes zero (excludes 1×). A pooled regression finds the same gap already present at the start of this catalog's window, 2023-05: open models priced 7.83× below closed ones even then, CI also excluding zero.

But the question backlog #97 actually asks — has that gap moved — comes back a clean null. The interaction term in log₁₀(price) ~ years + is_open + years·is_open, HC3-robust: +0.0790 log₁₀ points/yr, 95% CI [-0.1114, +0.2695] — contains zero, p=0.42. An independent check agrees: splitting the catalog at its own median release date, the gap in the early half is 4.79× (CI [2.65×, 11.18×], excludes zero) and in the late half 5.87× (CI [3.03×, 11.71×], excludes zero) — both intervals overlap heavily; neither half's gap is distinguishable from the other's. Within each group separately, price isn't detectably trending either direction: open-weight +0.0496 log₁₀/yr (CI [-0.0587, +0.1578], contains zero), closed -0.0295 (CI [-0.1629, +0.1040], contains zero). Pooled across both groups, ignoring the open/closed split entirely, the whole-catalog price trend also contains zero: +0.0674 log₁₀/yr, CI [-0.0334, +0.1683], R²=0.005.

Read together: this is not a story about a closing or widening gap. It's a story about a gap that has simply never moved, over three and a half years in which the underlying technology changed enormously. Open-weight models have always undercut closed ones by roughly five to six times on this metric, within the noise of a bootstrap confidence interval, at every point this catalog can check.

Limits, stated plainly. This is not quality-adjusted — it says nothing about whether a closed dollar buys more capability than an open dollar, only that it costs more per token moved. “Release date” is OpenRouter's own listing timestamp, not each lab's original announcement date, and a model can be re-listed when OpenRouter adds a new routing path for it. The sample is OpenRouter's own catalog, not every model either community has ever shipped — models no provider routes through OpenRouter don't appear here at all. And open-weight is a binary (hosted on Hugging Face or not), not a measure of how open a model's license, training data, or paper actually is.