Skip to content
</>

Cheapest LLM APIs for Indie Developers

LLM API pricing compared for indie workloads — where token costs, free tiers, and rate limits decide which API your side project can afford.

Top picks

Which tool should you pick?

Quick decision guide based on your situation.

Lowest absolute cost per token

DeepSeek API

Best free tier for hobby scale

Gemini Flash (Google AI Studio) or Groq free tier

One key, many models to compare

OpenRouter (aggregator)

Need speed above all

Groq (LPU inference on open models)

Production reliability and caching

Anthropic or OpenAI direct with prompt caching

Tool combinations that work together

How to combine multiple tools for a complete workflow.

Cheapest MVP loop

Gemini Flash free tier + edge cache

Free-tier inference with server-side response caching. Most prototypes never leave this tier.

Cost-optimized production

DeepSeek for bulk + Claude for hard cases

Route easy requests to the cheap model, escalate failures to the expensive one. Classic two-tier routing saves 80%+.

Multi-model A/B

OpenRouter + Vercel AI SDK

Swap models per request to compare quality/cost empirically before committing.

FAQ

What is the cheapest capable LLM API right now?

DeepSeek's API is the price floor for frontier-adjacent capability, typically an order of magnitude cheaper than top-tier models. Among US providers, Google's Gemini Flash models and GPT mini-tier models hit the best cents-per-million-token ratio for general work. For predictable costs, Groq offers fast open models (Llama-class) at low rates with generous free tiers.

Are free tiers enough for a side project?

Often yes. Gemini's free tier covers 15 requests/minute-scale usage, Groq's free tier handles thousands of requests/day of open-model inference, and Mistral offers a free experimentation tier. The catch: free tiers cap rate and may throttle at peak — design with caching and fallbacks so a 429 degrades gracefully instead of breaking the app.

How do I estimate my token costs?

Rule of thumb: 1 token ≈ 0.75 words of English. A chat reply of 300 words ≈ 400 output tokens. At $1/M output tokens, that's $0.0004 per reply — 2,500 replies for a dollar. Cost only bites with long contexts (which bill input repeatedly) or agent loops that re-send history. Cache system prompts and trim history to control both.

Should I proxy through an aggregator?

OpenRouter is the popular middleman: one API key, dozens of models, easy price comparison, and a small markup (~5%) for the convenience. Worth it while experimenting or serving multiple models; once you've standardized on one provider, going direct saves the margin and unlocks provider-specific features like prompt caching.

Get the next tool stack in your inbox

Weekly email for indie developers: new AI tools, curated stacks, and growth tips. No spam.

Free. Unsubscribe anytime.