The biggest price move on our model ledger this month belongs to DeepSeek. On 2026-09-28, DeepSeek V4 Pro 0813 went from $0.79 to $3.5 per million output tokens. A million tokens is roughly 750,000 words of model output, and the price of producing it rose about 4.4 times overnight.

One week later, on 2026-10-05, Reflection released Beam, an open-weight model it positions against Chinese models at lower compute cost, according to TechCrunch. The timing is convenient. Whether it is deliberate is Reflection's to say.

"Open-weight" means the trained model file is published, so a company can download it and run it on its own machines instead of renting it by the call. That matters this week for a plain reason. A rented model's price is set by someone else, and on 2026-09-28 someone else changed it. A downloaded model's price is your own electricity and your own chips. Nobody can reprice it on a Monday.

The claim deserves some skepticism, and here is what would have to be true for it to hold. "Lower compute cost" has to mean lower cost to serve the same quality of answer, measured on a named test such as SWE-bench or GPQA, on named hardware, per million tokens. A cheaper model that scores lower is just a cheaper model. The license also decides how open Beam is. Open-weight can mean "run it for anything" or "run it unless you are a competitor," and a buyer's lawyers will read that paragraph before any engineer reads the benchmark.

The comparison class is the part to watch. Our ledger holds 240 models, and it logged 28 price changes in the week to 2026-10-05. Prices on rented models move constantly. OpenAI's newest entry, GPT-6.1 Sol Pro, launched 2026-09-29 at $2 in and $10 out per million tokens. DeepSeek's new $3.5 output price now sits well under that, but it is no longer the rounding error it was at $0.79.

That is the shape of the choice in front of a finance team this quarter. One option is a rented Chinese model whose price just quadrupled. Another is a rented American model at $10 out. The third is a downloadable American model whose cost is yours to control, if Reflection's compute claim survives someone else's benchmark.

Anyone who budgeted on $0.79 per million output tokens has been budgeting on a number that expired on 2026-09-28.