What an AI Model Really Costs: Token Price, Token Consumption, Waiting Time

An AI vendor's price list answers a question nobody has. It states a price per million tokens — but an invoice is not paid in tokens, it is paid in completed tasks. And how many tokens a model needs for the same task varies by more than a factor of three.

Three kinds of cost, not one

Understanding what a model costs means looking at three figures together:

• The token price — what the provider charges per million input and output tokens. The only number in every price list, and the least informative one.
• Token consumption — how many tokens the model produces for the same task. Reasoning models bill their thinking steps too, even though those never appear in the answer.
• Time — how long until the first character, and how fast afterwards. This figure appears on no invoice and is still often the largest cost block, because a paid human is waiting in front of it.

Artificial Analysis measures all three on the same evaluation: cost of a full index run, tokens generated, output speed and time to first token.

The fourth column alone breaks price-list logic: Gemini 3.6 Flash needs 59 million output tokens for the same evaluation, DeepSeek V4-Flash 210 million — a factor of 3.6. Both score 50 on the index. Judged on token price alone DeepSeek looks 27 times cheaper; on the actual run the saving is tenfold, because the extra consumption eats part of it.

The Kimi K3 case: cheaper per token, but not half price

Kimi K3 costs 40 percent less per output token than Claude Opus 5 — $15 instead of $25. Less of that reaches the invoice, because Kimi burns 30 percent more tokens for the same work: 130 million against 100 million. What remains is a 36 percent saving. The discount survives, but it shrinks — and it was never half.

The rest is paid in time. Kimi K3 produces 34.9 tokens per second, Claude Opus 5 manages 54.9. More tokens at a lower rate means that for the same workload Kimi runs a good two times as long.

The reversal: inside an agent, the fast model wins

This is where simple rules of thumb break down, because there is a second time figure: time to first token. On reasoning models at high effort it covers the entire thinking phase before a single character appears. And there the models are worlds apart: Kimi K3 answers after 2.9 seconds, Claude Opus 5 after 85.6, Claude Fable 5 after 144.

In an agent with many small steps that value applies at every single tool call. A typical run of twenty steps with 800 output tokens each:

The result contradicts the section above it — and both are true. For one long generation Claude Opus 5 finishes sooner than Kimi K3 (13.6 against 19.1 minutes for 40,000 tokens). For twenty short steps the picture flips completely: Claude's thinking phase applies twenty times and adds up to more than half an hour, while GLM-5.2 is done after two minutes.

The lesson is not a model recommendation but a measurement rule: know the shape of your own workload before comparing models. Few long answers and many short steps are two different calculations.

Three workloads, worked through

The following calculations are our own, with assumptions stated. They use the list prices from the table above and assume the same token volume per item for every model — the real extra consumption of reasoning models comes on top.

Document extraction: 10,000 invoices a month

Assumption: 4,000 input tokens per document (the text of an average invoice), 800 output tokens per result. That makes 40 million input and 8 million output tokens a month.

Between the most and the least expensive option lie 792 dollars a month — on a task whose output can be checked mechanically (totals, tax rates, supplier numbers). That is exactly where a cheap model belongs: the error surfaces in validation, not at the customer.

Interactive assistance: 30 staff, 20 requests a day

Assumption: 12,600 requests a month at 2,000 input and 600 output tokens each. Pure token cost ranges from $5.65 (DeepSeek V4-Flash) through $14.11 (GPT-5.6 Luna) to $352.80 (GPT-5.6 Sol).

Now the time. The lead time to the first token applies to every single request:

* our own calculation, €50 fully loaded cost per working hour, 12,600 requests a month, highest effort level

The difference in token cost is $347. The difference in paid waiting time is over €21,000. That is the real reason effort level is not a detail in interactive applications: putting a reasoning model on "max" behind a chat a human operates burns working time on a scale that settles any token debate.

The agent run

Twenty steps of 800 tokens cost between half a cent (DeepSeek) and 80 cents (Fable 5) in tokens — negligible. What matters is the runtime from the chart above: 2.0 minutes against 51.5. For an agent that runs ten times a day, that is the difference between a tool people use and one they work around.

Self-hosting as the fourth option

For data that must not leave the building, the comparison is no longer between two APIs. Our own inference box with one RTX 5090 costs roughly €155 a month fully costed: about €111 depreciation (€4,000 of hardware over three years) and about €44 of power (240 watts average, €0.25 per kilowatt hour). On top of that comes maintenance effort that resists a single number.

That undercuts every model from Kimi K3 upwards (€222 and more) in the extraction workload above — and at the same time sits a factor of 21 above DeepSeek V4-Flash. The conclusion is uncomfortable but clear: self-hosting pays off against expensive APIs, not against cheap ones. Justify it on cost alone and you stand on weak ground. Justify it on data sovereignty and you are right — with cost parity against the upper model classes thrown in.

Two levers that appear in no comparison

Prompt caching shifts the maths considerably as soon as the same context is sent repeatedly — the normal case for agents, since the system prompt and tool descriptions go along with every step. DeepSeek charges $0.0028 instead of $0.14 per million tokens for a cache hit, one fiftieth. Build prompts so the stable part comes first and the input side costs a fraction.

Time-dependent pricing is arriving right now. DeepSeek has announced that during peak windows of 09:00–12:00 and 14:00–18:00 Beijing time — 03:00–06:00 and 08:00–12:00 in central Europe — every billing item doubles (price list). For batch runs: schedule them into the cheap hours and halve the bill. OpenAI goes the other way and sells speed as a surcharge — Fast Mode for GPT-5.6 Sol accelerates processing by up to 2.5× at double the price.

A grid for choosing

Batch work with nobody waiting — document extraction, bulk classification, translation runs. Only cost per workload matters, and verifiable output allows a cheap model. DeepSeek V4-Flash costs one fifty-third of Claude Opus 5, eleven index points behind.

Interactive work with a human in front of it. Waiting time dominates the bill by a factor of 60. Short lead time beats any index point here — or lower the effort level, which shortens exactly that lead time.

Agentic workflows with many steps. The decisive value is time to first token, not output speed. Models with a short lead time win even when they trail on the index.

Few, long, difficult generations. Here — and practically only here — the frontier model earns its price: higher index scores, less rework, and the long thinking phase applies once instead of twenty times.

Sensitive data. Self-hosting, for reasons of data sovereignty. Cost parity against the upper model classes follows by itself.

Our take

For our own workflows we run a gateway that routes requests by task rather than by brand: classification and extraction to cheap fast models, hard cases to a frontier model, sensitive data to local models on our own hardware. That split is exactly why the price collapse reaches us at all — it only works if the architecture stays swappable.

What we have stopped doing is comparing models by price list. The number that counts is the price per completed task — and next to it, the time somebody spends waiting for it. Neither appears in any quote.

Further reading

• Artificial Analysis — model leaderboard with cost and speed
• Claude Opus 5 — detailed measurements
• Kimi K3 — detailed measurements
• DeepSeek V4-Flash — detailed measurements
• Gemini 3.6 Flash — detailed measurements
• DeepSeek — official price list with cache and peak-hour rules
• The AI price collapse of 2026 — what the providers cut