How AI inference pricing works, and why a cheap token is not a cheap answer
Input, cached input, cache writes and output are four different prices, reasoning bills as output, and nobody publishes tokens per task.
Every large model vendor sells you at least four different prices at once, and the number on the headline rate card is the cheapest of them. Input tokens, cached input, cache writes and output tokens are billed separately, reasoning tokens bill at the output rate, and how many of each your task burns is decided by the model rather than by you.
That last clause is the whole problem. A rate card gives a price per unit. What you buy is an answer, and almost nobody publishes units per answer.
The four labs charge four different shapes of price
List prices per million tokens, from each vendor's own pricing page, as of 11 September 2026.
| Model | Input | Cached input | Output | Cheapest tier |
|---|---|---|---|---|
| GPT-6 Astra (short context) | $10.00 | $1.00 | $50.00 | $5.00 / $25.00 batch |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | $5.00 / $25.00 batch |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | $0.375 / $1.875 batch |
| DeepSeek V4.1-Flash (off-peak) | $0.15 | $0.003 | $0.60 | already the floor |
Read down the output column and the spread is 83 to one. Read the cached-input column and it widens to 333 to one, because a cache hit is where the providers have chosen to compete hardest.
Look at the two $10 models. Astra and Fable 5.1 are identical on input and output and differ by four times on cached input, because Anthropic prices a cache hit on Fable 5.1 at 0.025 times base input while every other Claude model uses the usual tenth. For a long-running agent replaying the same system prompt and the same repository, that column is most of the bill.
Output costs five times input, and reasoning is output
Every card above charges roughly five times as much for a token the model writes as for one it reads. That ratio is why the input price is the number vendors put in the tweet.
Reasoning tokens land on the expensive side. OpenAI's own guide is blunt about it: "While reasoning tokens are not visible via the API, they still occupy space in the model's context window and are billed as output tokens." Google says the same thing by putting the words in the label — its rate card line reads "Output price (including thinking tokens)". So a model that thinks harder costs more at an unchanged price, and the vendors say so out loud. Google, announcing Gemini 3.8 Flash on 2 September 2026, wrote that "3.8 Flash works harder" and that "at times, the model might use more tokens to maximize performance, especially at higher effort levels", which we covered in Gemini 3.8 Flash keeps its price and warns it will spend more tokens.
A per-token price that holds while per-task tokens rise is a price rise the rate card cannot show.
Caching is the largest discount on the page, and it has an entry fee
Anthropic charges 1.25 times base input to write a five-minute cache entry and two times base input for the one-hour version, then 10 percent of base input to read it (2.5 percent on Fable 5.1 and Mythos 5.1). OpenAI publishes a cache-write line too, at $12.50 against $10.00 base for Astra.
So caching pays only if you read more than you write. Rewrite the beginning of your prompt on every call and you have bought the surcharge without the discount. Anthropic's docs walk through exactly that failure, where a timestamp near the top of the prompt changes the prefix hash and "you pay for a fresh cache write on every request and never get a read". Put the volatile part last. It is the cheapest optimisation in this article.
Batch and off-peak discounts sell you patience
Anthropic, OpenAI and Google all take 50 percent off for asynchronous batch work, and Anthropic is explicit that most batches finish "in less than 1 hour" while a batch expires at 24. DeepSeek does the same trade by clock instead of by queue. Its pricing page says off-peak rates are half of peak, and that "peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)".
We counted the hours, because nobody else seems to have. That is seven peak hours a day across five weekdays, 35 of the week's 168. DeepSeek bills half price for 79 percent of the week, which makes the $0.15 headline the ordinary price and $0.30 the exception — the point of DeepSeek's V4.1-Flash costs $0.15 per million tokens for four-fifths of the week.
Ollama inverted the same trick. Its peak window runs 12:00 to 18:00 UTC on weekdays, 30 hours of 168, and doubles the price inside it. Eighteen percent of the week costs double instead of 21 percent costing half. The arithmetic differs from DeepSeek's by more than the marketing does.
Per-seat billing is becoming per-token billing
On 31 August 2026 Ollama moved its paid plans onto metered tokens, writing that Pro, Max and Team "now use transparent per-token pricing" with a monthly credit pool, and giving a reason worth keeping: customers said GPU-time billing "was difficult to predict, especially as open models have grown much larger (Kimi K3 has 2.8 trillion parameters)". Pro is $20 a month including $60 of usage, Max $100 including $300, Team $500 including $1,000 shared. Unused credit does not roll over. We wrote that up in Ollama put a weekday afternoon surcharge on open models.
Watch which direction your vendor is moving. A seat is a hedge the vendor sells you against your own variance, and vendors stop selling hedges when the variance gets expensive.
The token price is not a proxy for what the answer costs to produce
Vals AI published an audit on 3 September 2026 that priced this directly. It found that "DeepSeek V4 Pro, for instance, is nearly 15x cheaper per output token than Kimi K3 yet has roughly the same environmental footprint", that cost per task "fares no better as a predictor", and that the whole relationship is "further distorted by market dynamics, as model providers may offer tokens at steep discounts while trying to capture market share". Harvey, post-training its own model, put the builder's version in one line: "cost is a function of both token prices and tokens used."
The fair objection is that for plenty of work the rate card is fine. A classification job with a fixed prompt and a twenty-token answer has a stable token profile, and for that shape of work price per million really is price per task. The gap opens on agentic work, where effort settings, tool loops and reasoning budgets move the token count by an order of magnitude while the price per token sits still.
Gateways add a final layer. Straitly launched on 9 September 2026 advertising "5% CASHBACK ON CLOSED-SOURCE · 10% ON OPEN-SOURCE", which is a router handing back a tenth of the open-model spread and still trading. What do you think that spread is? Our reading of it is in A gateway paying ten percent back on open models has published its own margin.
How to actually price a workload
Measure tokens per completed task on your own traffic, at your own effort setting, then multiply. Four numbers do the work: cached input, fresh input, output including reasoning, and the share of calls that hit the cache. If you cannot produce those four, you do not have a cost estimate, you have a rate card.
Everything else in this article is a modifier on that number. Batch halves it if you can wait. Off-peak halves it if you can schedule. Caching cuts the input side by ninety percent or more if your prefix is stable, and raises it by a quarter if it is not.
What would change the answer
Three things, and one of them is scheduled. Gemini 3.8 Flash's introductory rate expires on 31 December 2026 and doubles to $1.50 and $7.50 on 1 January 2027, which will compress the table above from 83 to one down to about 42 to one on output. GPT-5.6 Sol's promotional pricing runs "at least through November 21, 2026". And if a lab ever publishes median tokens per task on a named benchmark alongside its rate card, this entire article collapses into one number and we will be glad to rewrite it.
More in our inference pricing coverage.
