Why start with vendor sites, not aggregators

Aggregator free credits usually come with expiry dates and routing limits, while a model vendor's own free tier gives you a stable console, a clear rate-limit table, and a predictable refresh cadence. In 2026, second-tier vendors (Mistral, Cohere, AI21, Groq, Cerebras) often offer more usable free volume than the two headline providers, and many don't require a card. Treat them as your daily supply layer and keep OpenAI/Gemini as a capability fallback.

Steps

Step 1: Build a vendor × capability list

Decide what kind of tokens you need: chat completion, embeddings, multimodal, or low-latency inference. List before you register.

Step 2: Register in no-card-first order

  1. Mistral: console offers a free trial tier; create an API key right after signup, usually no card required. Good for chat and embeddings.
  2. Cohere: trial keys with lower rate limits, enough for prototypes and batch embedding tests.
  3. Groq: low-latency inference; generate a key in the console, good for streaming and real-time validation.
  4. Cerebras: free tier open to developers, useful for high-throughput tests on large models.
  5. AI21: trial credits, good for long-text and summarization comparisons.
  6. Only then consider OpenAI / Gemini: higher barriers and more restrictions; use as fallback, not primary.

Step 3: Do only three things per account

  • Create a well-named key (e.g. proj-embed-2026), not the default one;
  • Run one minimal request via curl or the official SDK to confirm response shape and billing;
  • Find the Usage page and record used / limit / refresh time.

Step 4: Understand the refresh cycle

Most free tiers limit both requests-per-minute and monthly token totals. Put refresh times into your calendar or script comments so you don't burn out early in the month.

Step 5: Build a unified call layer

Use LiteLLM or a thin wrapper to map each key to a provider and route by task: embeddings to Cohere/Mistral, low latency to Groq, long text to AI21. When one vendor rate-limits, fall back automatically.

Caveats

  • Rate limits before quotas: the most common failure is RPM, not exhausted quota. Implement backoff retries.
  • Data terms: some free tiers allow requests to be used for service improvement; keep sensitive data off free tiers.
  • Key leakage: free keys are still credentials; keep them in env vars, never in repos.
  • Policy changes: vendors change free-tier terms anytime; screenshot the terms right after signup.

Use cases

  • Solo developers prototyping without a card;
  • Small teams needing a backup inference path not tied to one vendor;
  • Model comparison work that requires quick switching across vendors.