LLM API free tiers and default rate limits
Updated
On 2026-11-14, 6 of 14 LLM API providers we checked let a new account call the API free: Google Gemini API, Groq, Cerebras, OpenRouter, Cohere and BroadNet Inference. 2 state a dollar credit. Only Google Gemini API says free-tier data may improve its products. Entry limits run from 5 to 5,000 requests a minute.
Download CSV Download JSON Croissant metadata Compilation CC BY 4.0 · limits are the providers' published facts · checked 2026-11-14
BroadRouter is BroadNet's own site. BroadNet's row is included with the same fields as everyone else. This page makes no "best free tier" claim: the limits differ in kind, and the right one depends on your workload.
Key findings
- 6 of 14 providers let a new account call the API without paying, through a free tier, trial key, free models or a free credit: Google Gemini API, Groq, Cerebras, OpenRouter, Cohere and BroadNet Inference. 2 say the API needs payment (Mistral AI and Together AI), and 6 do not state a free option on the pages we read.
- 2 state a one-time free credit in dollars: Cerebras (USD 5) and BroadNet Inference (USD 5). Cerebras grants its credit only after a verified payment method is added; BroadNet (this site's owner) says no card is needed.
- 1 provider says free-tier data may be used to improve its products: Google Gemini API. 5 say API data is not used for training by default (OpenAI, Anthropic, Groq, Cerebras and BroadNet Inference). 8 make no statement about training on free-tier data on the pages we read (Mistral AI, Together AI, Fireworks AI, DeepInfra, DeepSeek, OpenRouter, Cohere and xAI); of those, Mistral has no free API and says paid API data may be used for training unless you opt out, and OpenRouter leaves it to each free model's provider.
- Published entry-level request limits range from 5 requests a minute (Cerebras, gpt-oss-120b) to 5,000 (OpenAI, Luna), across the 8 providers that publish an RPM figure. They are not comparable one to one: some apply per model, some per account or key, and the paid first tiers are much higher than free ones.
- 2 providers limit concurrent requests instead of requests per minute (DeepInfra and DeepSeek), and 3 publish no default numbers at all (Google Gemini API, Mistral AI and Together AI); Google shows Gemini free-tier limits only inside AI Studio.
- Card requirements are rarely stated on the documentation pages: 12 of 14 rows leave it unstated. Only Cerebras says a payment method is required for its free credit.
Which LLM APIs can you use for free?
| Provider | Free API access | Free credit | Card required | Trains on free-tier data | What is offered |
|---|---|---|---|---|---|
| OpenAI | Not stated | — | Not stated | No | The lowest usage tier is named Free: "User must be in an allowed geography", with a usage limit of 100 USD a month. The page publishes no per-model limits for it and no free credit amount. |
| Anthropic | Not stated | — | Not stated | No | No free tier or free credit stated on the rate-limits page. |
| Google Gemini API | Yes | — | Not stated | Yes | Free tier: many models show "Free of charge" on the pricing page, including Gemini 4 Hydrogen, Gemini 3.8 Flash and Gemini 3.5 Flash-Lite. |
| Mistral AI | No | — | Not stated | Not stated | Free mode includes a monthly usage budget for trying Vibe and Studio, but "It doesn't include API access." Creating API keys requires billing activation and pay as you go. |
| Groq | Yes | — | Not stated | No | Free plan with per-model limits; the Developer plan raises them. |
| Together AI | No | — | Not stated | Not stated | "Together AI doesn't currently offer free trials." A minimum credit purchase of 15 USD is required to use the platform. |
| Fireworks AI | Not stated | — | Not stated | Not stated | No free credit mentioned on the rate-limits page. |
| DeepInfra | Not stated | — | Not stated | Not stated | No free credit or free tier mentioned on the rate-limits page. |
| Cerebras | Yes | USD 5 | Yes | No | 5 USD in Free Trial credits after adding a verified payment method; they expire 30 days after they are granted. |
| DeepSeek | Not stated | — | Not stated | Not stated | No free tier or free credit mentioned on the rate-limit page. |
| OpenRouter | Yes | — | Not stated | Not stated | Free model variants (IDs ending in :free). "All new users receive a small free allowance to test out OpenRouter." |
| Cohere | Yes | — | Not stated | Not stated | Trial (evaluation) keys are "free but limited in usage": 1,000 API calls a month. |
| xAI | Not stated | — | Not stated | Not stated | No free credit or data-sharing programme mentioned on the rate-limits page. |
| BroadNet Inference | Yes | USD 5 | No | No | 5 USD starting credit for pay-as-you-go and hourly keys, no card needed. It works on BroadNet Gemma 4 VL 26B, BroadNet Transcribe Arabic and BroadNet Transcribe; the other models need a top-up first. |
Models that providers list at a zero price, model by model, are on free LLM API models.
What are the default rate limits for a new account?
Limits for the tier named in each row, for the model or scope named. Blank means the provider does not publish that figure. Do not compare columns across rows without reading the scope: per model, per account and per key limits behave differently.
| Provider | Tier | Applies to | RPM | TPM | RPD | Other | Source |
|---|---|---|---|---|---|---|---|
| OpenAI | Build (first paid tier: 5 USD in total credit purchases) | Luna | 5,000 | 2,000,000 | — | Build tier. Astra, Sol and Terra: 5,000 RPM and 1,000,000 TPM. No requests-per-day limits listed. | developers.openai.com · developers.openai.com (2026-11-14) |
| Anthropic | Start (monthly spend cap 500 USD) | Claude Haiku 5.5 | 4,000 | 2,000,000 | — | Start tier; TPM shown is input tokens per minute (ITPM, uncached only), with 400,000 output tokens per minute. New organizations may start in an Evaluation tier with lower limits. | platform.claude.com · privacy.claude.com (2026-11-14) |
| Google Gemini API | Free tier (Tier 1 needs a linked, active billing account) | — | — | — | — | Per-model free-tier limits are shown only in Google AI Studio: "Rate limits depend on a variety of factors (such as your usage tier) and can be viewed in Google AI Studio." The one model with published free-tier numbers is Gemini Decision 1 (200 RPM, 50,000 TPM, 10,000 RPD). | aistudio.google.com · aistudio.google.com (2026-11-14) |
| Mistral AI | Pay as you go | — | — | — | — | Limits are shown in the Admin Panel (API › Limits), not published as numbers on the page. | docs.mistral.ai · help.mistral.ai (2026-11-14) |
| Groq | Free plan | openai/gpt-oss-120b | 30 | 8,000 | 1,000 | Free plan, also 200,000 tokens a day. Whisper models: 20 RPM and 2,000 requests a day. | console.groq.com · console.groq.com · console.groq.com (2026-11-14) |
| Together AI | Minimum credit purchase of 15 USD | — | — | — | — | The rate-limits page gives no tier numbers for serverless models. | docs.together.ai · docs.together.ai (2026-11-14) |
| Fireworks AI | No payment method or no credits | Account-wide | 10 | — | — | 10 RPM without a payment method or credits; up to 6,000 RPM with a payment method and active credits. | docs.fireworks.ai (2026-11-14) |
| DeepInfra | Default account limit | Per public model | — | — | — | 200 concurrent requests. "every account has a default limit of 200 concurrent requests per model"; no fixed RPM or TPM. | docs.deepinfra.com (2026-11-14) |
| Cerebras | Free Trial | gpt-oss-120b | 5 | 30,000 | — | Free Trial: 30K uncached TPM (90K total TPM) and 1M tokens a day. qwen-3.8-27b: 300 RPM. | inference-docs.cerebras.ai · cerebras.ai · cerebras.ai (2026-11-14) |
| DeepSeek | Default account limit | deepseek-flash | — | — | — | 2,500 concurrent requests. Concurrency limit per account: 2,500 for deepseek-flash, 500 for deepseek-v4-pro; expansion on request at no extra cost. | api-docs.deepseek.com (2026-11-14) |
| OpenRouter | Free models, under 10 credits purchased | :free models | 20 | — | 50 | 20 RPM; 50 requests a day with under 10 credits purchased, 1,000 a day with at least 10. | openrouter.ai · openrouter.ai · openrouter.ai (2026-11-14) |
| Cohere | Trial key | Command A (Chat) | 20 | — | — | Trial: 20 requests a minute on Chat models and 1,000 calls a month; production keys 500 RPM on Command A. | docs.cohere.com (2026-11-14) |
| xAI | Tier 0 (0 USD cumulative spend, the default) | grok-4.7 | — | 50,000,000 | — | 150 requests per second. Tier 0: 150 requests per second and 50M TPM on grok-4.7; 37 RPS and 10M TPM on grok-4.3. "Your per-second limit is derived from your per-minute request budget (RPM / 48, with a minimum of 2)." | docs.x.ai (2026-11-14) |
| BroadNet Inference | Pay-as-you-go key | Per pay-as-you-go key | 30 | — | — | 4 concurrent requests. 30 requests a minute and 4 requests in flight per pay-as-you-go key; no tokens-per-minute limit is published for them. Hourly keys: 120 RPM, 1 stream, 300,000 TPM. | inference.broadnet.ai · inference.broadnet.ai · inference.broadnet.ai · broadnet.ai (2026-11-14) |
Which free tiers use your data for training?
- OpenAI: No. "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models" unless you opt in.
- Anthropic: No. "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models."
- Google Gemini API: Yes. Pricing page row "Used to improve our products": Yes on the Free Tier, No on the Paid Tier.
- Mistral AI: Not stated. No free API tier to apply it to. For the paid API, Mistral's help center says input and output data "may be included in Mistral's model training programs" and that you can turn this off with the Admin panel's Privacy toggle "to prevent API calls and related data from being used to improve Mistral's services"; it does not state the toggle's default.
- Groq: No. Services Agreement: "Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer." The agreement also covers fee-free services. Data page: "By default, Groq does not retain customer data for inference requests."
- Together AI: Not stated. No statement found on the pages we read.
- Fireworks AI: Not stated. No statement found on the pages we read.
- DeepInfra: Not stated. No statement found on the pages we read.
- Cerebras: No. Terms of Service: using Service Content to provide the Service "does not grant Cerebras the right to use Service Content for the purpose of training or fine-tuning models." Privacy policy: "we do not retain inputs and outputs associated with our training, inference, and chatbot Services".
- DeepSeek: Not stated. No statement found on the pages we read.
- OpenRouter: Not stated. Depends on the provider behind each model. OpenRouter routes to providers that "may train on your data (according to their own policies)" only if your privacy settings allow it, with "separate settings for paid and free models"; it does not state the default for free models.
- Cohere: Not stated. No statement found on the pages we read.
- xAI: Not stated. No statement found on the pages we read.
- BroadNet Inference: No. BroadNet's privacy page: "We do not use your inference data to train, fine-tune or evaluate models." Prompts, audio and answers are not retained.
If a free tier may use your prompts, do not send customer data, secrets or anything under a confidentiality agreement through it. Read more on zero data retention.
Method
- Each value was read with AI assistance from the provider's own documentation, pricing or privacy page on 2026-11-14; the JSON keeps the source URL and check date for every row. Third-party aggregators were used only to find pages, never as a source.
- Limits are the defaults a new account gets on the lowest tier the provider documents with numbers: the free tier where one exists, otherwise the first paid tier. They are copied in the provider's own unit (per minute, per second, per day, or concurrent requests) for the model named in the row.
- A field is empty ("not stated") when the provider's pages did not say it. We did not infer card requirements from sign-up flows, and we did not convert per-second or concurrency limits into requests per minute.
- "Trains on free-tier data" records only an explicit statement on the provider's own page, terms or privacy policy. For OpenAI, Anthropic, Groq and Cerebras the statement covers all API data, free or paid.
- BroadNet's credit comes from pricing.json (synced 2026-11-14) and its limits from the FACTS block in src/data/product.ts, both read from BroadNet's own pages.
Left out:
- Mistral AI, Free-tier rate limit numbers: Third-party sites quote figures, but Mistral's own docs show limits only inside the Admin Panel.
- Google Gemini API, Free-tier RPM/TPM/RPD for Flash and Pro models: Google's docs say these are shown only in AI Studio.
- All, Card requirements not stated on the pages we read: Left empty rather than inferred from sign-up flows.
Update frequency: re-read with AI assistance; the check date changes when it is. BroadNet's row is recomputed on each build. Limits: published defaults only; your account page may show different limits. Code: src/lib/dataset-free-tiers.ts.
What do the CSV columns mean?
| Column | Meaning |
|---|---|
id | Row identifier. |
provider | Company selling the API. |
free_api_access | yes if API calls are possible without paying (free tier, trial key, free models or a free credit); no if the provider says it needs payment; empty if not stated. |
free_credit_usd | One-time free credit in USD, where stated. |
free_offer | The free offer in the provider's words, shortened. |
entry_tier | Tier the limits in this row belong to. |
limit_model | Model or scope the limits apply to. |
rpm | Requests per minute. |
tpm | Tokens per minute (input tokens per minute for Anthropic). |
rpd | Requests per day. |
rps | Requests per second, where the provider publishes per-second limits. |
concurrency | Concurrent requests, where the limit is on requests in flight. |
limit_note | Other limits and caveats. |
card_required | yes, no, or empty if not stated. |
trains_on_data | yes if the provider says free-tier data may be used to improve its models, no if it says it is not used, empty if not stated. |
data_note | The data-use statement, quoted or summarised. |
source_urls | Provider pages the row was read from, separated by " ". |
checked | Date the sources were read (YYYY-MM-DD). |
How to cite this dataset
Our compilation (the selection, normalised fields, counts, tables and method): CC BY 4.0. Free to use, share and adapt with attribution and a link to this page.
Each limit, credit and data-use statement is a fact published by the provider named in its row, at the source URL given per row; we do not license those pages to you. Credit the provider as the source.
Plain text
BroadRouter (2026). LLM API free tiers and default rate limits for new accounts (version 2026-11-14) [Data set]. BroadNet Technologies. https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits/BibTeX
@misc{broadrouter_llm_free_tiers,
author = {BroadRouter},
title = {LLM API free tiers and default rate limits for new accounts},
year = {2026},
version = {2026-11-14},
publisher = {BroadNet Technologies},
howpublished = {\url{https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits/}},
note = {Data set, data as of 2026-11-14},
}Questions
Which LLM APIs have a free tier?
On 2026-11-14, Google Gemini API, Groq, Cerebras, OpenRouter, Cohere and BroadNet Inference offered some way to call the API without paying. The terms differ: daily request caps, trial keys, free model variants or a one-time credit. See the table for each.
Which LLM API free tiers can use my data for training?
Of the providers we checked, Google Gemini API says free-tier data is "used to improve our products" (No on its paid tier). OpenAI, Anthropic, Groq, Cerebras and BroadNet Inference say API data is not used for training by default. On OpenRouter it depends on the provider behind each free model and on your privacy settings; the others make no free-tier statement on the pages we read.
Do I need a credit card for a free LLM API?
Not always, but most documentation pages do not say. BroadNet says its USD 5 credit needs no card. Cerebras requires a verified payment method before it grants trial credit. Fireworks caps accounts with no payment method or no credits at 10 requests a minute.
What are the default rate limits for a new LLM API account?
They vary by two orders of magnitude. The lowest published entry limit is 5 requests a minute (Cerebras); first paid tiers at OpenAI and Anthropic allow thousands. Some providers cap concurrent requests instead (DeepInfra 200 per model, DeepSeek 2,500 on deepseek-flash). Your account page shows your exact limits.
Is OpenRouter's free tier limited?
Yes. Free model variants (IDs ending in :free) are limited to 20 requests a minute and 50 requests a day, or 1,000 a day once you have bought at least 10 credits, per OpenRouter's limits page.
Can I use this data?
Yes. Our compilation and counts are CC BY 4.0; cite them as shown under "How to cite". The limits and terms are each provider's published facts: credit the provider, and check its page before you rely on a limit, because tiers change often.
Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity
Sources
- OpenAI: developers.openai.com, checked 2026-11-14
- Anthropic: platform.claude.com, checked 2026-11-14
- Google Gemini API: aistudio.google.com, checked 2026-11-14
- Mistral AI: docs.mistral.ai, checked 2026-11-14
- Groq: console.groq.com, checked 2026-11-14
- Together AI: docs.together.ai, checked 2026-11-14
- Fireworks AI: docs.fireworks.ai, checked 2026-11-14
- DeepInfra: docs.deepinfra.com, checked 2026-11-14
- Cerebras: inference-docs.cerebras.ai, checked 2026-11-14
- DeepSeek: api-docs.deepseek.com, checked 2026-11-14
- OpenRouter: openrouter.ai, checked 2026-11-14
- Cohere: docs.cohere.com, checked 2026-11-14
- xAI: docs.x.ai, checked 2026-11-14
- BroadNet Inference pricing.json, synced 2026-11-14