DeepSeek R1 and the Falling Price of Intelligence

On September 12, 2024, OpenAI released a model that thinks before it answers, and Plus users could send it 30 messages a week.²

On January 20, 2025, DeepSeek released a reasoning model under an MIT License, with an API price a small fraction of the other model's.¹ Four months. That is how fast the price of this kind of intelligence moved.

If you run a small business, you do not need to follow the lab drama. You need to understand what the price curve does to your decisions. My answer: model cost is collapsing, and process design is the scarce skill.

What did DeepSeek release on January 20, 2025?

A reasoning model called R1, plus a technical report and smaller versions of the model.

What DeepSeek says R1 can do

DeepSeek describes R1 as having performance on par with OpenAI-o1 across math, code and reasoning tasks. It says the model and technical report are fully open source, with code and models released under the MIT License, which permits commercial use and distillation. It also released six smaller models distilled from R1, and says the 32B and 70B versions are on par with OpenAI-o1-mini.¹

I cannot verify those benchmark claims myself, and neither can you from a press release. That is the first lesson. A claim is a reason to run a test, not a reason to change your stack.

Why "reasoning model" matters

Reasoning models spend more compute thinking through a problem before they answer. OpenAI's o1-preview was the first widely available example (I wrote about what it changes for operators): OpenAI said it was trained to spend more time thinking, with strengths in science, coding and math, and noted that for many common cases GPT-4o would still be more capable.² That is useful for planning, analysis and multi-step tasks. For simple drafting you may not need it.

Which of your tasks actually need one?

Most small business work does not. Rewriting an email, summarizing a call or pulling action items rarely needs a model that thinks for a long time. The tasks that might are the ones with several steps and real tradeoffs (the same split as in agents versus workflows): comparing offers, checking a pricing model for holes, reviewing an ad report for what is actually driving the result, or planning a sequence of emails against a goal.

If you cannot name a task of that kind in your week, the price drop is interesting but not urgent. If you can, it is a good time to rerun the math on what it is worth to hand that task to a model and review the output.

Plays here, from YouTube

How big is the price gap?

Large enough to change what is worth automating.

DeepSeek's page lists $0.14 for cached input, $0.55 for uncached input and $2.19 for output, per million tokens.¹ OpenAI's pricing page, as captured on January 25, 2025, listed o1 at $15 per million input tokens and $60 per million output tokens.³ Divide them and you get roughly 27 times on both input and output. That is my arithmetic from two published price lists, not an independent benchmark of cost for the same job.

The caveats on the price

Price per token is not price per task. Reasoning models can use a lot of tokens thinking before they answer, so cost depends on how many a task burns. Prices also move, and different hosts charge differently. Do the math on your own workload.

Should your business switch to DeepSeek?

Probably not on the strength of a headline. Probably yes to a careful test.

A sane way to evaluate

• Pick three real tasks you do repeatedly, such as summarizing call notes, drafting a proposal outline or analyzing an ad report.
• Run them on two models. Judge the output blind if you can.
• Track cost and time per task, not per token.
• Check where your data goes. Read the provider's terms before you paste anything about a client.
• Keep a human on anything that reaches a buyer.

An open license also means the model can be run by other providers or inside your own environment, which is a data-control option the closed models do not offer in the same way. Whether that is practical for a one-person business is a separate question.

What does cheap intelligence make scarce?

Stanford's 2024 AI Index estimated about $78 million of compute to train GPT-4 and $191 million for Gemini Ultra.⁴ Building frontier models is still enormously expensive. But the cost to use intelligence is falling even while the cost to build it stays huge, and users are the ones who benefit.

When thinking is cheap, four things get expensive by comparison.

• Process design. Which steps, in what order, with what inputs.
• Clean inputs. Your offer, your voice, your customer data, organized so a model can use them (start with getting your data ready for AI).
• Review. A person who can tell when the output is wrong.
• Accountability. Someone who owns the result, especially when it reaches a customer.

In a business with no staff, the thing that holds my work together is not which model I use that month. It is a process I understand well enough to hand off in pieces.

How do you design for a price that keeps falling?

The operators who win this cycle will not be the ones who guessed the cheapest model. They will be the ones whose workflows were clean enough to swap it in on a Tuesday. If you want help designing yours, reach out.

Keep building.Ekai