The Cheapest LLM API, by What You Actually Run

The cheapest LLM API is not a model you can name once and keep. Price is two numbers, dollars per million input tokens and dollars per million output tokens, and models set those two so differently that the ranking shifts with the job. On most workloads we priced gpt-oss-120b comes first, but once a job writes much more than it reads, Mistral Nemo overtakes it, and further down the list the gaps between models open and close the same way. All prices below are read from the OpenRouter model list on 2026-11-05, the same source behind our model price table; to price your own workload across every model at once, use the LLM cost calculator.

Here are eleven representative models, cheapest first, with a monthly bill for one worked workload: a support chatbot running 50,000 requests a month at about 1,200 input and 350 output tokens each.

The spread on one ordinary workload is from about a dollar a month to three thousand. That gap is the reason to pick deliberately, and why a headline "cheapest" number tells you almost nothing until you attach it to your own volume.

Why there is no single cheapest API

Every provider bills the same two events: the tokens you send in the prompt and the tokens the model writes back. Output tokens are generated one at a time, while the prompt is read in one parallel pass, so output usually costs several times more than input. GPT-5 shows the shape plainly, at $1.25 per million in and $10 out, an eightfold step. Claude Haiku 5.5 is gentler, $0.10 in and $0.50 out. Mistral Nemo is almost flat, $0.029 and $0.03.

Your bill is those two prices multiplied by how much your workload reads and writes, then added. A job that reads a lot and writes little lives on the input price. A job that writes a lot lives on the output price. The two rankings are not the same list, which is why a calculator that only knows a model's headline price cannot answer the question, and why pricing the actual workload is the only reliable way to compare. A unit price is no forecast either: four questions break most automation ROI numbers, and a model bill is one more input to them.

The cheapest API for a chatbot

A chatbot writes. On the worked example above, 60 million input tokens and 17.5 million output tokens a month, the output price carries most of the bill. The open-weight gpt-oss-120b wins outright at about $1.38, because both its numbers are tiny. Among the small closed models from the big labs, GPT-5 Nano comes to about $10 and Claude Haiku 5.5 to about $15, both decided more by their output price than their input. A frontier model on the same traffic is a different order of spend: GPT-5 near $250, Claude Opus 5.5 near $590.

The lesson for a conversational product is to look at the output price first and the input price second, then test the two or three cheapest on your own transcripts before committing. For support replies a small model is often enough; test it on your own data. The difference between $15 and $295 a month adds up every month the bot runs.

The cheapest API for classification and extraction

Classification flips the arithmetic. Take a pass that sends 250 million input tokens a month and returns 10 million, roughly 500,000 short documents in and a label out. Now the input price decides almost everything. gpt-oss-120b again wins, near $3.16. Two closed models show the input price at work: GPT-5 Nano, at $0.05 per million input, comes to about $16.50, and Claude Haiku 5.5, at $0.10, to about $30, nearly double. On the chatbot the same pair came to $10 and $15, a smaller gap, because their output prices are closer. For a real change of leader, turn the job around: at 250 million input and 750 million output tokens a month, Mistral Nemo comes to about $30 and gpt-oss-120b to about $34. The models did not change. The workload did.

This bites when a pilot moves to production. If a team picks the model that was cheapest for the demo's conversational traffic and keeps it for a batch job where the input price now dominates, the bill can land well above the estimate. The fix is to price the production workload, not the demo; once it runs, the three numbers boards ask for show whether it paid off.

The cheapest API for coding and long-context work

Coding and retrieval-augmented work send big prompts and write moderate output. On a workload of 30,000 requests a month at 6,000 input and 1,500 output tokens, both prices matter. A small model such as Claude Haiku 5.5 comes to about $40 and GPT-5 Nano to about $27, while GPT-5 runs near $675 and Claude Sonnet 5.5 near $810. The gap is real, but this is also the workload where quality is most often worth paying for, because a wrong answer to a hard coding task costs a human hour, not a fraction of a cent.

Context length is the other axis here. The cheapest models on the list stop near 131K tokens. Million-token windows start around tenth place on this workload, with Qwen3.5-Flash and DeepSeek V4 Flash, and include Gemini 2.5 Flash Lite and Claude Haiku 5.5, so a long document or a large codebase chunk fits without splitting. That is worth checking on the model price table before you choose, because a cheap model with a short context can cost more in engineering workarounds than a dearer one that holds the whole input.

When the expensive model is the cheap choice

Price per token is not price per useful answer. A model that fails one task in five and needs a retry, a human correction or a fallback call is not cheap, however low its sticker. On reasoning, hard code and anything customer-facing, the frontier models are worth their price where a wrong answer is costly; test that on your own tasks. The rule is to route by difficulty: the cheapest model that holds the quality for the easy majority of calls, an expensive model reserved for the minority that needs it. Decide this per step in the brief for an AI integration, next to the process, its owner and its metric. The savings are in that decision, not in shaving a tenth of a cent off the unit price of the wrong model.

How to read a provider's price before you trust it

Three things move the real bill past the two headline numbers. First, cached input: on the list GPT-5 Nano charges $0.005 per million for cached tokens against $0.05 for fresh ones, which matters a lot for a long system prompt repeated on every call. Second, batch pricing: the batch versions of GPT-5 Nano and Claude Haiku 5.5 cost half the standard rate, which suits classification and summarisation but not a live chat. Third, check whether the price of the model you pick changes with prompt length, because a long-context workload could then cost more per token than the table shows. None of these change the arithmetic, but they change the absolute figure, so read the provider's own pricing page for the model you settle on. For the wider market of AI products around these models, see our sweep of 10,066 tools.

The same model can carry more than one price

For open-weight models the question has a second layer, because the weights are public and several providers host the same model at different rates. The gpt-oss and Llama families, Qwen, Mistral's open releases and DeepSeek are offered by several hosts, so compare the hosts of the model you pick before you settle on one. An aggregator such as OpenRouter can route a request to another host of the same model, which saves wiring a separate account for each. Our table takes the list price for each model from that same aggregated source, which keeps the comparison in one unit, but when you settle on an open-weight model it is worth checking who hosts it and at what rate, because the host can still change the bill noticeably. A closed model such as GPT-5 or Claude has one list price, and this layer does not apply.

How to choose

Count three things before you open any pricing page: your monthly request volume, the average input tokens per request, and the average output tokens per request. Put those into the calculator and read the ranking for your own numbers, not for a demo. Take the two or three cheapest that look plausible on quality and test them on your real data. Then wire in one cheaper fallback, because prices and models move through the year and a workload tied to a single model is a workload exposed to its next price change.

The cheapest LLM API is the one with the lowest bill at your own ratio of input to output tokens. Price the workload you actually run, on both numbers, before you pick.