What distillation actually costs

There is no single number, but the cost structure is predictable and the dominant term is not the one most people expect.

The three lines

1. Teacher queries. You pay per token to generate the corpus. This scales with how much capability you are transferring and is usually the largest line.

2. Training compute. Fine-tuning the student. Real, but for a small student on a prepared corpus it is often a smaller number than the query bill.

3. Evaluation and iteration. Consistently underestimated. You will not get the prompt distribution right first time, and each revision means regenerating part of the corpus.

Working the arithmetic

The query cost is straightforward to estimate before committing:

cost = examples x avg_tokens_per_example x price_per_token

The sensitivity is entirely in the first term, and it varies by orders of magnitude with scope:

• One narrow task: thousands of examples. Often a few hundred dollars.
• A domain: hundreds of thousands. Meaningful but ordinary project spend.
• Broad general capability: millions to tens of millions of exchanges. This is a different category of budget, and it is the scale the 2026 disclosures described.

Reasoning traces cost more per example than answers, because they are longer. If you are doing reasoning distillation, assume a multiple on average token count.

Why the comparison to pretraining is the whole point

Set against training a comparable model from scratch, distillation is cheap by orders of magnitude. Pretraining a frontier-class model requires a compute budget available to very few organisations. Distilling from one requires an API account and a training run.

That asymmetry is the entire reason distillation became a geopolitical argument rather than an engineering footnote. Export controls raise the cost of the first path. They do not touch the second. See What Is a Frontier Model?

Before you spend anything

1. Try quantization first. It is nearly free and often sufficient. See What Is Quantization?
2. Try a smaller off-the-shelf model. It may already be good enough for your task.
3. Pilot with a small corpus. Ten thousand examples will tell you whether the transfer works before you commit to a million.
4. Only then scale. Most failed distillation projects failed at step three and found out after step four.

Common questions