Choosing a free AI API: tiers, trials and compute credits
Choose by task, payment requirements, quota cycle and data terms. Separate free chat interfaces from hosted inference and compute.
Before copying a key, identify what is free: a chat interface, model weights, hosted inference or compute for your own deployment. These are different products.
Four questions first
- What task do you need: chat, transcription, embeddings or self-hosted inference?
- Is payment-method verification acceptable?
- Is this a one-off evaluation or a recurring daily job?
- Will prompts contain private user data? Review data terms before selecting a model.
Compare like with like
A recurring free tier, an expiring trial balance and a compute allowance cannot be compared by their headline number. Cloudflare uses Neurons; Modal's compute credit excludes Shared Endpoints; Cerebras currently offers a time-limited trial after payment verification.
Build your own decision sheet
Record the model, supported task, quota unit, reset cycle, payment requirements, data policy, overage behavior and review date. Unknown does not mean unlimited.
Use the same small set of public tasks for each candidate. Record whether it meets your needs and why failures happen. Do not call a documentation comparison a hands-on benchmark.
Budget the whole workflow
An agent task may make many requests and repeat its context. Estimate requests, input size and retries for a complete task rather than one response. Cap runtime, output length and retry count. A paid fallback needs its own explicit budget.
References
OpenRouter limits, Workers AI pricing, Modal pricing, Cerebras trial. This is a selection framework, not a performance ranking.