The short answer
A free AI API can help you evaluate models or run a low-volume prototype, but “free” does not mean unlimited, private, or production-guaranteed. Compare the specific model’s quota and reset period, data-use terms, and paid rates—and check what happens when you reach a limit. As of November 8, 2026, Google’s Gemini API, GroqCloud, and OpenRouter offer different kinds of free access, but their quotas are not directly interchangeable.
This guide focuses on hosted generative language-model APIs, not consumer chatbot subscriptions. Offers and prices can change; confirm the current terms for your account, model, and endpoint before relying on them.
Free tier, trial credit, or local model?
- Free tier: Ongoing API access up to a quota. Limits can apply to requests, tokens, models, or features. Reaching a limit can block requests; moving to paid access may require a separate billing step.
- Trial credit: A temporary or spend-limited balance. It can help with testing, but is not the same as an ongoing free allowance. Check expiry, signup requirements, and what happens when it runs out.
- Local inference: Running a model on your own hardware can avoid hosted per-token API charges, but hardware, electricity, setup, and maintenance still cost money. It is not a free hosted API.
Free AI API comparison
| API | Free access and model | Limits and reset | Signup, payment, and what happens at the limit | Other decision points |
|---|---|---|---|---|
| Google Gemini API | Free access is limited to certain models. Google’s pricing page does not establish one universally eligible model identifier in the current information cited here. Check its pricing and model list before choosing an ID. | Limits depend on model and project; check your account’s rate limits. Google says daily request quotas reset at midnight Pacific Time. | New accounts start on the Free tier, though some projects may need verification. Moving to Paid requires linking billing and prepaying at least $5. The cited billing documentation does not specify what happens for every model and project when a free quota is exhausted; check your project’s limits and billing settings before enabling paid access. See billing details. | Google says Free-tier content is used to improve its products, while Paid-tier content is not used for that purpose. Confirm applicable terms and regional conditions for your use. |
| GroqCloud | Free-plan access is model-specific. The limits page lists openai/gpt-oss-20b at 30 requests/minute, 1,000 requests/day, 8,000 tokens/minute, and 200,000 tokens/day. |
Limits apply at the organization level, and any request or token limit may be reached first. The cited documentation does not establish this model’s daily reset time; check your organization’s limits page. | The cited documentation does not establish whether payment details are required at signup or give a model-specific explanation of behavior at daily quota exhaustion. Rate-limit information and retry headers are documented; check current account limits before depending on the free allowance. | The model page lists paid input and output rates; see the cost example below and Groq’s model details. Groq says inference data is not retained by default, subject to documented exceptions; see data handling. |
| OpenRouter | The Free plan includes access to free models, including the openrouter/free router, which selects from available free models. |
The pricing page lists 50 requests per day on Free. The documented counter resets at midnight UTC. Model availability and limits can vary. | The Free plan is available without a paid plan, but its pricing page does not specify whether payment details are required to sign up. Exceeding the free request cap results in rate limiting. A fallback to a paid model can incur charges, so use only free models in a fallback list if you need to keep requests free. | The pricing page lists no contractual SLA on Free. On Standard, it lists a 5.5% fee when purchasing credits; inference is billed at the provider’s listed model rate. |
These limits measure different things: Gemini quotas vary by model and project, Groq’s example has request and token limits, and OpenRouter’s stated free cap is a request count. They are not directly comparable. A workload can stay below a daily cap and still hit a per-minute limit.
Which one should you try first?
- Try Groq if
openai/gpt-oss-20bfits your task and its current organization-level quotas work for your traffic. If you handle confidential data, review Groq’s retention terms and controls before sending it. - Try OpenRouter if comparing available models through one API is more useful than sticking to one provider. Treat the selected model and its backend provider as part of the decision: routing does not make their limits or data policies identical.
- Try Gemini only after confirming that a specific model ID is currently listed as Free for your project and that its Free-tier data-use terms suit your task. The pricing page cited here establishes limited access to certain models, but not a specific eligible model ID to recommend.
A direct provider API may offer that provider’s own features and quota. A router can make model comparisons easier but adds its own limits, fees, and terms. Neither option is automatically cheaper or safer; decide based on the specific model, workload, and policies you need.
What will it cost after free access?
API cost usually depends on model, input tokens, output tokens, and sometimes other billable items. Estimate input and output separately. These provider-listed examples are not a claim that the models have equivalent quality or capabilities:
- Groq
openai/gpt-oss-20b: $0.075 per million input tokens and $0.30 per million output tokens, as listed on its model page. - OpenRouter
openai/gpt-oss-20b: $0.004 per million input tokens and $0.017 per million output tokens, as listed on its model page. Confirm the selected model and serving route when you make a request. - No Gemini estimate is included: without a confirmed Free-eligible model ID, a model-specific paid-rate estimate would not be a like-for-like comparison. Check the current Gemini pricing page for the model you select.
Worked estimate: Suppose an application makes 10,000 requests per month, averaging 1,000 input tokens and 200 output tokens per request. Assuming all tokens are billed at the listed rates, that output totals all billable output tokens, and no other charges apply:
| Example endpoint | Input calculation | Output calculation | Estimated model usage |
|---|---|---|---|
Groq openai/gpt-oss-20b |
10 million × $0.075 = $0.75 | 2 million × $0.30 = $0.60 | $1.35/month |
OpenRouter openai/gpt-oss-20b |
10 million × $0.004 = $0.04 | 2 million × $0.017 = $0.034 | $0.074/month, before any credit-purchase fee |
These are paid-rate estimates, not automatic bills triggered by exceeding a free quota. OpenRouter says its Standard-plan fee is charged when credits are purchased, rather than on individual inference requests. The minimum credit purchase or minimum fee is not established here, so the usage subtotal is not the same as the cash needed to start paid use. Check current pricing and purchase terms before budgeting. For any provider, check billing settings before enabling paid use. A monthly average can also hide a busy day or a burst that exceeds a per-minute limit.
Business-use risks to check
Data use and retention: Google’s pricing page distinguishes Free-tier content, which it says is used to improve products, from Paid-tier content, which it says is not used for that purpose. Confirm the terms and regional provisions that apply to your project. Groq says inference requests are not retained by default, with documented exceptions; see its data policy. Groq also documents an optional Zero Data Retention control. Confirm its scope directly before relying on it. OpenRouter says each provider has its own data-handling policy; review its provider-logging guidance and the policy for the backend that will receive your request. A router is not one uniform data-retention policy.
Commercial and model permissions: Read both the API provider’s terms and the license or terms for the specific model. This comparison does not establish commercial-use permission for every provider-and-model combination. A free API offer alone is not business approval. If a service routes among providers, check which provider will receive the request.
Reliability and limits: Free access can be useful for testing, but quotas and model availability can constrain customer-facing service. OpenRouter lists no contractual SLA on its Free plan. Confirm that the specific paid service and contract offer any availability or response-time commitments you require; do not infer them from a free tier.
Budget controls: Before linking billing or buying credits, set a budget, monitor token usage, and decide what the app should do when a request is limited: queue work, return a clear error, or switch to a vetted fallback. Check the fallback’s costs, privacy terms, and limits too—especially whether it can route to a paid model.
Bottom line
For an evaluation, start with the API whose specific model, data terms, and quota unit fit your test—not the one with the largest-looking number. Before business use, verify current model-level limits, signup and billing conditions, paid input and output rates, retention and training terms, model permissions, and any reliability commitment you need. Free access can lower the cost of trying an API; it does not replace that review.