MMOOZZGG.com

Neural network API

Free LLM API: limits, terms and the price for your data

7 min read17 sources

A free LLM API is offered by several providers as of 16 October 2026, but each limits the volume, and some take payment in data. The Gemini free tier gives requests to Google to improve its products, DeepSeek announces no free tier, and OpenAI and Gemini are officially unavailable from Russia. Below are the limits from official pages and the cases where the free option ends up costing more than a paid one.

In short: free options as of the date

As of 16 October 2026, the official pages of the providers give this picture:

  • Gemini API: many models have a free tier, including Flash, Flash-Lite and Gemini 2.5 Pro; free tier requests go to improving Google’s products.
  • OpenAI API: the documentation has a Free tier without buying credits, but only for allowed countries; Russia is not on the list.
  • Claude API: no free tier; there are Start, Build and Scale tiers with a monthly spending cap; Anthropic gives new users a small free credit to test the API.
  • DeepSeek API: the pricing page has no free tier; requests are paid from a topped-up balance.
  • Qwen in Alibaba Cloud Model Studio: new users get a quota, usually 1 million tokens per model for 90 days.
  • Mistral: Free mode without a bank card, with limited limits; requests in this mode may go to training.
  • OpenRouter: models marked :free, 20 requests a minute and 50 a day without buying credits.
  • Open models DeepSeek R1 and Qwen 3: the weights are free under their licenses; you pay for your own hardware.

Free tiers of providers

Gemini API. On Google’s pricing page as of 16 October 2026, a free tier (Free of charge) is available for Gemini 3.8 Flash, 3.6 Flash, the Flash-Lite models, Gemini 2.5 Pro and 2.5 Flash, and for the open Gemma 4. Gemini 3.1 Pro Preview and the image and video models are not in the free tier.

Google does not publish exact free tier limits on the page; AI Studio shows them for your project. Limits are counted in requests per minute, tokens per minute and requests per day, and the daily quota resets at midnight Pacific time. The limit applies to the project, not the key, so a second key adds no quota.

OpenAI API. The Free tier is available in allowed countries without buying credits, and the first paid tier, Build, opens after you buy 5 dollars of credits. Russia is not on the OpenAI API list of countries as of 16 October 2026.

DeepSeek API. The pricing page has no free tier. The price is per million tokens, separately for input with a cache hit, input without cache and output; off-peak rates are half as much. If the account has a granted balance, it is used before the topped-up one.

Qwen. Alibaba Cloud gives new Model Studio users a quota, usually 1 million tokens for each model, for 90 days and only in the Singapore region. By default, after the quota runs out, requests continue and are charged at the rate; the Free Quota Only setting turns on a stop.

Mistral and OpenRouter. In Free mode, Mistral opens the API without a bank card and calls the limits restricted, without figures. For models marked :free, OpenRouter gives 20 requests a minute and 50 a day; after you buy 10 credits, the daily limit grows to 1,000 requests.

Paid plans are easier to compare elsewhere: token prices are in the API price comparison, and the cloud API calculator works out the monthly bill.

How you pay: requests go to training

The Gemini API terms say it directly: on the free tier, Google uses requests and answers to improve its products, and human reviewers may read them. Google asks you not to send confidential or personal data there. On the paid tier, requests are not used to improve products, but they are stored for a limited time.

Mistral says something similar about its Free mode: requests and answers may go to model training, and you can opt out at any time (as of 16 October 2026).

Aggregators are more complex. OpenRouter passes requests to different providers, each with its own retention and training policy. You can block providers that train models separately for paid and free models, and by OpenRouter’s own words this setting does not change its own policy for handling requests.

For a company in Russia, the law adds more. Under 152-FZ, before a cross-border transfer of personal data the operator files a notice with Roskomnadzor, and data of Russian citizens is collected and stored in databases in Russia. The questions to ask any provider about data are collected on the page about API security.

Open models: free on your own hardware

The second free path does not require giving away data. On 20 January 2025 DeepSeek released the R1 model with open weights under the MIT license. On 29 April 2025 the Qwen 3 family came out with open weights under the Apache 2.0 license.

The weights download for free, the model runs on your computer or server, and requests do not go to a provider. You pay for hardware and electricity, not for tokens, and you set the request limit yourself.

For a trial, a program on a work computer is enough; a team needs a server with a graphics card. How Ollama, LM Studio and vLLM differ and which to choose for your load is covered separately.

What your own server costs and when it pays back compared with the cloud is worked out by the local AI cost calculator.

When free stops working

A free tier is meant for a trial. Four signs that it is time to move to a paid plan or your own model:

When volume grows, a dedicated API on the open models DeepSeek and Qwen makes sense: the OpenAI format, your requests do not go to training, and you see the logs and the request processing yourself.

  • Volume: 50 requests a day for OpenRouter’s free models and Gemini’s daily quotas run out on the first real flow of leads or documents.
  • Data: requests now contain clients, contracts, personal data. The Gemini free tier with human reviewers and Mistral’s Free mode with training do not fit them.
  • Predictability: on its limits page, Google writes that the stated limits are not guaranteed in advance. For a service that answers clients, this is a risk of downtime.
  • The bill: by default, the Alibaba Cloud quota turns into paid requests without a stop unless you turn on Free Quota Only.

Questions and answers

Is there a free LLM API? Yes. As of 16 October 2026, the Gemini API has a free tier, Qwen in Alibaba Cloud has a quota for new users, and OpenRouter has free models.

Each has strict limits, and on the free tier Gemini uses requests to improve its products. Gemini and OpenAI are officially unavailable from Russia.

Can a business use a free API? For a trial on anonymized data, yes. For client data, no: Google asks you not to send personal or confidential data to the free tier, and 152-FZ, Russia’s personal data law, requires a notice to Roskomnadzor, the data protection regulator, before personal data is transferred abroad.

A production load needs a paid plan or an open model on your own servers.

Sources and dates

Contact us