The cheapest LLM APIs in 2026: price per million tokens compared (OpenAI, Together, OpenRouter, GPU Flow)

How is LLM API pricing calculated?

Almost every LLM API charges per million tokens, with two prices: one for input tokens (your instructions, the conversation so far and any documents you attach) and one for output tokens (what the model writes). Output is usually several times more expensive, because generating each token takes more GPU time than reading one.

Two more lines can move the bill. Cached input is a reduced price for the part of a prompt the provider has already seen, such as a long system prompt repeated on every request. And reasoning models usually bill their thinking as output tokens; Google, for one, writes it into its Gemini API price table. A model that thinks for 2,000 tokens before answering in 300 is billed for 2,300 output tokens.

What makes an LLM cheap for your workload?

To compare like for like, the tables below price two example requests. A support answer: about 500 input tokens and 300 output tokens. A retrieval (RAG) query over your documents: about 6,000 input tokens and 400 output tokens. Both are our assumptions, chosen to show how the order changes when input dominates.

Four things decide which model is cheapest for a given job.

How much does the OpenAI API cost per million tokens?

OpenAI publishes several service tiers. These are the standard prices, in US dollars, for its GPT-5 family; for gpt-5.5 and gpt-5.4 they are the rates for prompts under 272K tokens, which cost more above that size.

OpenAI standard-tier prices, in US dollars, checked on 9 November 2026. Source: developers.openai.com/api/docs/pricing

The nano and mini models are the ones to compare with small open models. The OpenAI models are proprietary: you cannot download them or run them on hardware of your own.

What does Together AI charge for open models?

Together AI serves open-weight models from several labs. A selection of its serverless chat prices, in US dollars per million tokens. Several of these models also have a lower cached-input price on the same page.

Together AI serverless inference prices, in US dollars, checked on 9 November 2026. Source: www.together.ai/pricing

Together AI lists Gemma 4 and Qwen3.8 27B only in its fine-tuning table on that page, not among the serverless inference prices, so they are left out here.

What do models cost on OpenRouter?

OpenRouter is a router: one API in front of many providers, each with its own price for the same model. The figures below are the price OpenRouter lists for each model in its public model list, in US dollars per million tokens.

OpenRouter model list and pricing page, in US dollars, checked on 9 November 2026. Sources: openrouter.ai/api/v1/models, openrouter.ai/pricing

Each model page also lists the price of every provider that serves it, some above and some below the figure here, and where a request lands depends on your routing settings. On the Standard plan OpenRouter adds a 5.5% platform fee, according to its pricing page, so the token price is not the whole bill.

What does GPU Flow charge per million tokens?

These are current Token Factory prices for every text and vision model in our catalogue, in euros with VAT included, read from the same source as our pricing page. The cached-input column is filled only where we bill a cache read.

Same model, different providers: how do prices compare?

Several open models are on more than one of these platforms. The table puts them side by side, each price in its own currency, with the cost of 1,000 RAG queries for scale.

Two cautions before you pick a winner from this table. The same model name is not always the same build: we serve GLM 5.3 in its NVFP4 quantised version, for example, and on a router each provider chooses its own format. And a lower number in dollars is not automatically a lower bill in euros once you add the exchange rate, card fees and VAT. On some rows the dollar figure is lower than ours, on others it is higher; check the rows for the models you actually plan to use.

How this works on GPU Flow

Token Factory is billed in euros, with VAT included, so the price in the table is the price on your invoice, without currency conversion; customers in the Canary Islands, Ceuta, Melilla or outside the EU, and EU businesses with a valid VAT number, pay without VAT. There is one rate for everyone, with no volume tiers and no monthly fee.

You load credit and it goes down as you use it. When it runs out, the API answers with a 402 error until you top up; nothing is charged to your card automatically, so a bug in a loop cannot run up a bill beyond your balance. The API is OpenAI-compatible, so moving a workload over is a change of base URL, key and model name, and what you send is never used to train any model.

What else should you check besides the token price?

1. 1Currency and tax: a dollar price is converted by your bank on each payment, and may or may not include VAT for your country.
2. 2Platform and card fees: routers and some providers charge a percentage on top of the credit you buy.
3. 3Minimum purchase: some providers require a first top-up before the first request, and that money is committed.
4. 4Rate limits: the cheapest model is no use if your account can only send a handful of requests a minute.
5. 5Reasoning length: a model with a low output price that thinks for thousands of tokens can cost more per answer than a pricier one that does not.

Related

• Free LLM API credits: six providers’ free tiers compared, and what our welcome credit buysThe free tiers of OpenAI, Gemini, Groq, Together AI, OpenRouter and Mistral, checked on their own pages, and what the GPU Flow welcome credit pays for.
• An OpenAI API alternative: GPU Flow vs OpenAIPrice per million tokens, audio, migration and when to stay on OpenAI, with their prices checked on their own site.