Adding AI to your app: models, free tiers and gateways
What AI model calls cost per month, which free tiers publish their limits, which use your data, and when an AI gateway is worth adding.
· 6 min read
In short: small, fast models cost a few dollars a month for a thousand users. The larger models cost 20 times as much or more. Free tiers are good for building and testing, but most stop at low request limits, and some use your prompts to improve their products. All prices below were checked in October 2026.
How AI APIs bill
You call a large language model (an LLM) from your code through an API, and pay per token. A token is a small chunk of text, often part of a word. Every request has:
- Input tokens: your prompt, instructions and any context you send.
- Output tokens: the model's reply. These usually cost several times more than input.
Prices are quoted per million tokens. The stack builder assumes 10 AI requests per user a month, each with 1,500 input tokens and 500 output tokens. At 1,000 users, that is 15 million input tokens and 5 million output tokens.
What models cost
Per million tokens, and what the builder's 1,000-user example costs each month:
| Model | Input | Output | 1,000 users |
|---|---|---|---|
| Mistral Ministral 3 3B | $0.10 | $0.10 | $2.00 |
| Groq GPT OSS 20B | $0.075 | $0.30 | $2.63 |
| OpenAI GPT-6 Luna | $0.10 | $0.50 | $4.00 |
| Anthropic Claude Haiku 5.5 | $0.10 | $0.50 | $4.00 |
| Mistral Small 4 | $0.15 | $0.60 | $5.25 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $17.00 |
| Gemini 3.8 Flash | $0.75 | $3.75 | $30.00 |
| OpenAI GPT-6.1 Sol | $2 | $10 | $80.00 |
| Claude Sonnet 5.5 | $2 | $10 | $80.00 |
| Claude Opus 5.5 | $4 | $20 | $160.00 |
Cost grows in a straight line with use. At 10,000 users, multiply by ten.
Some details change the bill:
- Claude Haiku 5.5's price is for prompts up to 100,000 tokens. Longer prompts cost $0.50 in and $2.50 out.
- Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text.
- Gemini's output prices include thinking tokens, the hidden reasoning a model writes before its answer. Reasoning-heavy replies cost more than the visible text suggests.
- Cached input is cheaper at OpenAI, Anthropic, Gemini and Mistral. Cached input is the part of a prompt the provider has seen recently and stored, such as the same instructions sent with every request.
- Batch requests are half price at Anthropic and Mistral, and on Gemini's Flash models. A batch request is sent in a group to be processed later, instead of answered right away.
Announced price changes
Google has announced price rises for some Gemini models:
- Gemini 4 Argon goes from $2 in and $10 out to $4 in and $20 out on December 1, 2026.
- Gemini 3.8 Flash and 3.6 Flash go from $0.75 in and $3.75 out to $1.50 in and $7.50 out on January 1, 2027.
That doubles the 1,000-user example for 3.8 Flash from $30 to $60 a month. The builder shows a badge on these models and uses today's price in its estimates.
Mistral Large 4 is listed at a sale price of $0.68 in and $2.09 out. Mistral's model page gives the original price as $1.36 in and $4.18 out, with no end date for the sale.
Free tiers and their limits
| Free tier | Published limits | Uses your data |
|---|---|---|
| Groq Free | 30 requests a minute and 1,000 a day, per model | Not recorded in our data |
| OpenRouter Free | 20 requests a minute and 50 a day, free models only | Depends on the model provider |
| Gemini free tier | Not published; shown only in Google AI Studio | Yes, to improve Google's products |
| Mistral Free mode | Not published; shown in your Mistral Admin Panel | Yes, for training by default; you can opt out |
Groq also limits tokens on Free: 8,000 a minute and 200,000 a day for these models, and you hit whichever limit comes first. The Developer tier (pay as you go, card required) raises requests to 1,000 a minute and 500,000 a day.
OpenRouter limits free use to its free models. Buying at least $10 of credits raises the cap to 1,000 requests a day. OpenRouter's FAQ says free models are usually not suitable for production. Some providers behind it may train on your prompts, and you can block those in your account settings.
Gemini's free tier covers 3.8 Flash, 3.6 Flash and 3.5 Flash-Lite. Google's pricing page says free-tier content is used to improve its products, and paid-tier content is not.
Mistral may use Free mode inputs and outputs to train its models by default. Free and paid users can both opt out.
A free tier bills no tokens, so the builder shows it at $0 until your traffic passes its limits. It checks daily limits against a busy day, assumed to be twice an average one. At 1,000 users, the busy day is about 667 requests: inside Groq's 1,000, but far over OpenRouter's 50. Gemini and Mistral are not offered as free plans in the builder, because their limits are not published.
AI gateways
An AI gateway sits between your app and the model providers. It gives you one API for many models, plus features like caching, fallbacks and spending limits.
- OpenRouter: one API for hundreds of models at the providers' prices. It charges a 5.5% fee when you buy credits, with a $0.80 minimum per card purchase, and credits expire after a year. Bringing your own provider key has no fee for the first $25,000 a month of list-price use.
- Vercel AI Gateway: no markup on tokens, and every Vercel team gets $5 of free credit a month for a subset of models. Paid use runs on credits you buy, and you pay the payment processing fee, which Vercel does not state. Vercel's Hobby plan is non-commercial, so a commercial product needs a Pro team.
- Cloudflare AI Gateway: caching, rate limits and analytics are free when you use your own provider keys. Paying for models through Cloudflare instead adds a 5% fee on credits.
None of these three lowers the token price, and OpenRouter and Cloudflare add a fee when you buy credits through them. A gateway is worth adding when you want one API for several models, fallbacks, caching or usage limits in one place.
What to pick
- Building and testing: Groq Free has the clearest published limits. Gemini's free tier is fine for experiments, but its content is used to improve Google's products, so keep private user data out of it.
- A first product on a budget: a small model such as GPT-6 Luna, Claude Haiku 5.5, Ministral 3 3B or GPT OSS 20B. Each costs a few dollars a month at 1,000 users.
- Quality matters more than price: test a mid-size model like Claude Sonnet 5.5 or GPT-6.1 Sol on your real prompts, and budget about $80 a month per 1,000 users at the builder's defaults.
- Several models, or you want to switch later: add a gateway. Cloudflare AI Gateway costs nothing with your own keys.
Price your app in the stack builder, or browse LLM APIs & Inference and AI Gateways & Sandboxes.