After DeepSeek: What Open-Source Models Mean for Enterprise AI Costs

On January 20, 2025, DeepSeek released its reasoning model R1: weights open-sourced under an MIT license that explicitly allows commercial use and distillation, benchmark results pitched against OpenAI's o1, and API output priced at ¥16 per million tokens (about $2.19). Within weeks the name had travelled from engineering forums to dinner tables.

It is now late October, nine months on. For an industry review the timing is better than it was in January — the noise has faded, and what remains can be seen plainly.

Pin the facts down first

Before the retrospective, the facts. R1 is a 671-billion-parameter mixture-of-experts (MoE) model that activates only about 37 billion parameters per inference pass, with a 128K context window. Launch API pricing: ¥1 per million input tokens on a cache hit, ¥4 on a miss, ¥16 per million output tokens — $0.14, $0.55 and $2.19 on the official dollar list. Alongside R1 came six distilled smaller models — distillation being, roughly, a large model teaching a smaller one — from 1.5B to 70B parameters, with the 32B and 70B versions benchmarked by DeepSeek against o1-mini. The license is MIT: commercial use, modification and distillation are all explicitly permitted — a detail that returns in every layer of impact below.

What got reset was the anchor, not one price

Before R1, "o1-class reasoning" was scarce and priced like a scarce good. R1 offered capability benchmarked against o1 at ¥16 per million output tokens, and the significance goes beyond one discount: it re-anchored what reasoning capability is supposed to cost. Every vendor pricing a model since has had to answer the same question — compared with what open source makes freely available, what exactly justifies the premium? Once an anchor moves, every quote in the market gets re-set against it.

Three layers of impact for businesses

Layer one — companies calling APIs: experimenting got cheap. For most smaller companies, AI means calling an API. With reasoning prices down, scenarios that never used to pencil out become worth a try: clause-by-clause contract comparison, structured summaries of long reports, bulk data cleaning and classification. These tasks lean on reasoning, and a single full run used to cost enough to kill the idea; now you can pilot small and scale on evidence. One caution: cheap unit prices invite ballooning usage, and the bill does not necessarily shrink — covered in full in how enterprises control AI costs.

Layer two — vendor negotiations: you hold a new card. When a capability existed only inside one or two closed models, procurement had no leverage and lock-in was the default ending. Open weights changed the structure of the conversation: "switch to self-hosted open source" now sits permanently on the table, even if you never exercise it. In system design, keep the model access layer thin and standard, and never couple business logic deeply to one vendor's private interface — not distrust, just ordinary procurement sense extended to AI.

Layer three — private deployment: the bar dropped for real. MIT licensing plus permitted distillation means a company may take the weights into its own server room; the six distilled models pull the hardware bar for "a good-enough private model" from cluster-grade down to single-machine-grade. For businesses with sensitive data, isolated intranets, or one fixed scenario at very high call volume, the route is broadly viable for the first time. It is no free lunch — operations, updates and evaluation land on you; where it fits and where it does not is explored in private deployment and small models.

The sober side: cheap models did not make cheap applications

The past nine months also showed the other face. We have watched companies green-light projects on the strength of "models are dirt cheap," only to find months later that the budget never lived in the model. Documents must be shaped into machine-usable form, processes taken apart and rewired, staff taught to actually use the thing and to check its output — none of that has ever had a price cut. The model is the cheapest component of an AI application, and R1 strengthened that truth rather than weakening it: the lower the unit price, the larger the share of total cost held by people and process. Reading "model price cut" as "AI project price cut" has been the most common miscalculation of the year.

How to choose models in a multi-model era

After R1, the right question is no longer "which model do we pick" but "which scenario gets which model": routine summaries and format conversion run fine on budget tiers; complex reasoning and high-stakes output get a reasoning model plus human review; sensitive data goes private. From an implementation standpoint, what we watch for is whether a company keeps its own evaluation habit — a small test set of twenty or thirty real tasks from your business, run against each new model, beats following launch events. Match models to scenarios; do not pledge allegiance to vendors. It is the most useful discipline of the open-source era.

Nine months later, what remains

Looking back, three things settled out of those noisy weeks: a lower price anchor, an open-source alternative that never leaves the table, and a private-deployment route with a far lower bar. All three point the same way — model capability is turning from a scarce good into infrastructure. For businesses, the competitive weight shifts further from "do you use AI" to "how well do you use it" — and the homework for using it well was never on the model's side.