The cheapest AI API is not the cheapest to run

Every comparison chart ranks AI APIs by price per million tokens. That number does not tell you what a task costs. A model priced at $0.14 per million input tokens can finish a job for more money than one priced at $3.00, because the cheap model may write ten times as many tokens getting there.

There is a second problem with those charts. The provider that anchors the bottom of nearly every one of them has announced that its prices are going up.

Why price per token misleads you

Price per token is the amount a provider charges for text going in and coming out, quoted per million tokens. It is easy to compare, which is why every pricing table uses it. It is also only half of a cost calculation.

The other half is how many tokens the model actually spends. That is not a property of the price list. It is a property of the model.

A reasoning model is a model that generates intermediate reasoning text before producing its final answer. You are billed for that intermediate text at the output rate, even though you never display it. The published research puts the overhead at a minimum of 3.11 times and a maximum of 14.78 times the token cost of a non-reasoning model on the same benchmark tasks (the arXiv study of reasoning-model token overhead).

The overhead is also badly distributed. That study found the overspend concentrates on easy problems, where the extra reasoning changes nothing. Some models burned over 900 tokens answering a trivial arithmetic question.

What the current prices actually are

Here is what the providers publish, in USD per million tokens.

Model	Input	Output	Cache hit input
DeepSeek V4 Flash	$0.14	$0.28	$0.0028
DeepSeek V4 Pro	$0.435	$0.87	$0.003625
Claude Haiku 4.5	$1.00	$5.00	$0.10
Claude Sonnet 5 (through August 31, 2026)	$2.00	$10.00	$0.20
Claude Sonnet 4.6	$3.00	$15.00	$0.30

DeepSeek figures come from the company's published rate card. Claude figures come from Anthropic's model pricing table.

The headline gap looks enormous. DeepSeek V4 Flash output is roughly 1/54th the price of Claude Sonnet 4.6 output. That ratio is what drives the "switch and save almost everything" advice circulating in cost guides.

Now apply the token multiplier. If the cheaper model is a reasoning model and spends 10 times the output tokens on your workload, the real ratio is closer to 1/5th, not 1/54th. Still cheaper. Not remotely the saving advertised.

If the workload is one where reasoning helps little, and the cheaper model retries or rambles, the gap can close entirely.

The number almost no chart shows you

Prompt caching is the largest single lever in that table, and it is not a model choice at all.

Prompt caching is a billing mechanism where a provider charges a reduced rate for input tokens it has already processed and stored. On DeepSeek V4 Flash, input drops from $0.14 to $0.0028 per million tokens on a cache hit. That is a 50x difference on the input side of every request.

Most real applications send a large, stable prefix on every call. A system prompt, tool definitions, retrieved documents, a conversation history. That prefix is exactly what caches well.

Two teams using the identical model can therefore see input bills that differ by an order of magnitude, based purely on whether their prompt structure is stable enough to hit cache. Restructuring a prompt so the fixed content sits at the front, unchanged between calls, is usually a bigger win than switching providers.

The size of that discount also varies by provider, which no pricing table shows. DeepSeek's cache hit rate of $0.0028 against $0.14 is a 50x reduction. Anthropic prices a cache read at 0.1x base input, a 10x reduction. A workload with a large stable prefix therefore shifts the provider comparison in DeepSeek's favour by roughly a further factor of five, on top of the headline gap. A workload that varies every prompt gets none of that.

This is also the reason agent workloads get expensive so fast. An agent re-sends its accumulated context on every step. Our earlier look at whether AI agents can run a business, where the agent lost $447, is a cost story as much as a competence story. Steps multiply tokens, and tokens multiply everything else.

The floor is moving

DeepSeek's pricing page currently states that the company plans "to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected."

That matters beyond DeepSeek's own customers. Its rates are the reference point that cheap-tier comparison charts are built around. Those charts still quote $0.14 and $0.28 as settled facts. The vendor has said in its own documentation that they are not.

DeepSeek is not alone. Claude Sonnet 5 carries introductory pricing of $2.00 input and $10.00 output per million tokens through August 31, 2026, after which standard pricing of $3.00 and $15.00 takes effect (Anthropic's introductory pricing note). That is a scheduled 50% increase with a published date on it.

If you are planning next quarter's inference budget against today's numbers, build in headroom at both ends of the market. A provider that announces a significant increase rarely delivers a small one.

The open-weights route is the usual hedge, and it is less available than it looks. Chinese labs have been slower to release weights than their announcements suggested, as the unfulfilled Qwen 3.8 Max open weights promise showed. Self-hosting also swaps a token bill for a GPU bill, which is only cheaper above fairly high, steady utilisation.

What to measure instead

Stop comparing price per million tokens. Measure cost per completed task. This is not a niche preference; it is how the benchmark industry already reports cost. Artificial Analysis defines cost per task as input, cached, and output token prices multiplied by tokens actually consumed, and states plainly that "models that produce longer answers or more reasoning tokens will have a higher cost per task, even at identical per-token prices" (Artificial Analysis benchmarking methodology).

The procedure for your own workload is short.

1. Take 50 real requests from your application, not benchmark prompts.
2. Run them through each candidate model.
3. Record total input tokens, total output tokens, and how many outputs were actually acceptable.
4. Compute the total dollar cost divided by the number of acceptable outputs.

That last division is the one people skip. A model that is half the price per token and fails twice as often is more expensive per usable result.

Two adjustments to run alongside it:

• Check your cache hit rate before you check anyone's price list. On DeepSeek's rates, a good hit rate is worth more than most model switches.
• Cap reasoning effort where your provider exposes that control. Much of the measured overhead is spent on easy inputs that never needed it, according to the benchmark analysis of reasoning token spend.
• Check whether batch processing fits. Anthropic discounts both input and output by 50% for asynchronous work, per its batch processing rates, which beats most model downgrades without changing model quality.

So which API should you use?

For most production workloads, the honest answer is that per-token price should be roughly the fourth thing you consider.

Choose on task success rate first. Then on cache behaviour with your prompt shape. Then on whether the provider's data terms fit your product. Then on price.

Budget-tier models earn their place in narrow, high-volume, low-ambiguity jobs. Classification, extraction, routing, short summarisation. These are tasks where reasoning overhead stays small and failure is cheap to detect. A published worked example puts roughly 3,700 tokens on an average support conversation, or about $37.00 per 10,000 tickets on a small model (Anthropic's support-agent cost walkthrough). That is the shape of workload where cheap models genuinely win.

Frontier models remain the better economics wherever a wrong answer costs real money, or where a retry loop can quietly triple the token count you budgeted for.

The bottom line

The cheapest AI API on a price list is a claim about one input to your bill, not about the bill. Reasoning overhead, cache hit rate, and failure rate each move the real number by more than the headline rate does.

Measure cost per completed task on your own traffic. Then treat the published rates as what they are right now: a floor the largest budget provider has already said it intends to raise.