DeepSeek made tokens cheap. Your workflow is still expensive.

The last week of January was loud. A lab in China shipped a reasoning model that made a lot of Western pricing decks look theatrical. Slack filled up with “so we can 10x the agents now?” We got three calls in four days that started with the word DeepSeek and ended with no description of the job.

Cheaper tokens are good. We like cheaper tokens. They are also a trap if you were already confusing a demo with a system.

The useful question is not “which model is winning this month”. It is “what does a wrong answer cost, and who notices”. That question was correct in 2023. It is just cheaper to ignore now, which is why more teams will.

The bill was never the hard part

Even at last year’s prices, most of the AI programmes we audited were not dying from OpenAI invoices. They were dying because nobody owned the 8% error case. A claims assistant that invented a policy clause. A support draft that promised a refund the company does not give. A “research” agent that cited a PDF it never saw.

When tokens get cheap, those failures get cheaper to produce at volume. That is not a win. That is a firehose aimed at your operations team.

We watched one team celebrate a 70% drop in model spend and then hire two people to clean up emails the system had sent. The ledger looked great if you only opened the inference tab.

Price the workflow, not the model

We put four numbers on the page before we pick a provider.

1. Cost per successful task, including retries, tools, and the human who still reviews it.
2. Time to a usable answer. A 40-second reasoning trace that a clerk waits on is slower than a 4-second retrieval with a form.
3. Cost of a wrong answer. If it sends money, changes a medical record, or emails a customer, this number dominates.
4. Concentration risk. Can you switch models without rewriting prompts that are secretly application logic?

DeepSeek, Gemini, Claude, GPT — pick based on those four, plus where the data is allowed to live. If your counsel will not let prompts leave India or the EU, the leaderboard is a hobby. Self-hosting a distilled model can be the grown-up answer even when a hosted API is “smarter” on a blog bench.

Prompts that encode business rules are the concentration risk people skip. “Never refund above 2,000 unless the tag is VIP” does not belong in a system prompt. It belongs in code, with a test. Models change. Your refund policy should not change because a provider shipped a new default temperature.

What we actually changed in builds that month

We started routing easy classification to small models and keeping the expensive reasoner for the 15% of cases that fail a confidence check. That pattern was correct before January. It is just less awkward to explain now that the cheap model is not an embarrassment.

Confidence is a messy word. We do not trust a model to rate its own certainty unless we have calibrated it on our set. More often the gate is mechanical: missing fields, low retrieval overlap, a tool error, a customer in a regulated cohort. Those gates are boring. They catch the cases that make the news.

We also stopped letting product managers paste a system prompt into a slide and call it a spec. The spec is the evaluation set. Fifty real examples from last quarter, labelled by the person who does the job. If you do not have those, you do not have an AI feature. You have a party trick.

• Log the retrieved context, the tool calls, and the final action. Not just the pretty answer.
• Put a human on anything irreversible. “Irreversible” includes emails. People forget that.
• Cap spend per tenant per day. Cheap tokens still cluster when a loop goes feral at 2 a.m.
• Keep a boring fallback: search, a form, a queue. The day the provider blips, the business still has to close.
• Version the prompt and the eval set together. A prompt change without a score is a deploy you cannot explain.

A note on “we will fine-tune”

Fine-tuning shows up in the same meetings as DeepSeek, usually as a way to feel in control. Sometimes it is right — classification on a stable label set, a house style, a language the base model keeps dropping. It is the wrong first move when your retrieval is bad or your tools are sloppy. You will bake the sloppiness in and then spend a month wondering why the new weights still invent policy numbers.

Residency is not a footnote

A lot of the January excitement assumed prompts can go anywhere. Plenty of the work we do cannot. Health, lending, payroll, anything with a government id. If the answer is “the data stays in India” or “in the EU”, you are choosing providers and VPCs before you choose a leaderboard. We have walked away from a cheaper model because the processing region was a shrug.

Write the constraint on the same page as the four numbers. If counsel has not spoken, you do not have an AI design. You have a prototype you will have to throw away in front of a customer.

Logs are residency too. If you send prompts to a US observability vendor, you have left India even if the model is in Mumbai. We have had to unwind that. It is not a fun conversation with counsel. Put the trace store on the same map as the model.

If January taught the market that intelligence is getting cheaper, good. Our job did not change. Make the workflow honest, then spend as little as you can on the model that survives contact with Tuesday.