Free LLM APIs for developers without unreliable shortcuts

Searching for a free LLM API usually turns up a list of providers and a column of limits. Before you build around one, check what makes the calls free, who shares the quota, and what happens when it runs out. An agent can reach those limits quickly, especially when several jobs use the same account.

What “free” actually costs you

Every provider prices “free” differently, and the differences matter more than the headline.

Gemini applies quotas per project, with daily requests resetting at midnight Pacific time. Its data terms vary with region and billing status: unpaid-service content may be used for product improvement, while paid-service data rules also apply to unpaid quota for users in the EEA, Switzerland, and UK. Making an API client available to users in those regions requires Paid Services. Check those conditions before sending sensitive data.

Groq measures requests and tokens per minute and per day at organization level. Its rate-limit table gives different ceilings even within GPT-OSS: Safeguard has lower request and token throughput than the chat models. Read the model row and your console limits before sizing an agent loop.

Match the free-access mechanism to your workload

These directory pages explain the differences without treating every offer as renewable free capacity:

• OpenRouter, Requesty, and Routeway expose selected free model IDs. Requesty shares its free-model quota across the organization; Routeway documents conflicting per-minute limits. Switching model IDs may leave the same cap in place.
• Mistral includes monthly usage, with exact organization/model limits in its Admin Panel. Cloudflare Workers AI meters its daily allocation in Neurons, whose token equivalent varies by model. Workers Free stops at the allocation; Workers Paid bills excess usage. SambaNova ties its documented Free Tier to an account with no payment method; its public plans and limit pages disagree about setup.
• Cerebras offers finite trial credits after a verified payment method, expiring after 30 days. It has no permanent free tier. NVIDIA NIM trial access is restricted to internal testing and evaluation under its trial terms.
• SiliconFlow has separate international and China accounts, endpoints, and access terms. OVHcloud distinguishes anonymous testing from authenticated paid inference. Nous Portal documents subscription access; a zero-price catalog row alone does not establish free-account entitlement.

Check the current account plan as well as the model price. Paid APIs with signup credits stop being cost-free when those credits expire or run out.

There is no reliably unlimited hosted free LLM API

Plan for metered hosted inference. The offers covered here all have access conditions or quotas. Treat any “no limit” claim as unverified until the provider documents what is metered and for how long. The cap may apply to requests, tokens, a time window, model access, or a shared resource such as an IP address. Spread your workload across independently usable quotas and decide how your application should handle a request when all eligible routes are exhausted.

How to select a provider

Providers change free tiers, quotas, and model catalogs. Ficelle’s provider comparison records a source-review date and links each provider’s own docs. Check these points before combining providers:

• Quotas. Check eligible models, regions, account conditions, and the request and token ceilings.
• Tool support. Check that the model can perform the tool calls your agent needs.
• Privacy. Check retention, training settings, regional terms, and the policies of upstream providers.
• Health checks. Catalogs list capabilities that models do not always deliver. Check the capability your application needs with a real request.
• Failover. What happens the moment your primary free model rate-limits, gets retired, or silently drops a feature it used to support.

Plan how your application handles a failed request: provider breakage is the silent killer of AI agents, and a single free key has no defense against it.

How Ficelle applies the framework

Ficelle is a local, OpenAI-compatible router. Its default catalog path requires zero-price evidence. Optional quota providers need configured access and explicit free-access guards; their presence in the provider pack does not prove your account can use them. Strict-zero routing keeps paid fallback disabled. When a model rate-limits, Ficelle can retry another eligible candidate. If none can serve the request, it returns an error. Keep quota-provider accounts within their free entitlement: local selection cannot prevent an upstream account from billing beyond its allocation.

It runs on your machine against a local endpoint, with credentials configured locally and requests sent directly to the selected provider rather than through a Ficelle-hosted proxy. Its routing code is source-available under the Business Source License 1.1.

None of this makes free models universally equal to paid ones. In Ficelle’s free-versus-paid benchmark, GPT-5 won 10 of 20 polished-content tasks, so long-form writing is one workload worth keeping on a paid model or with a human editor.

If you are a developer tired of rebuilding the same quota-and-failover logic by hand for every free LLM API you try, compare Ficelle’s plans and let a maintained pool do it instead.