Comparing LLM API Prices: How to Think About OpenAI vs Anthropic vs Google

"Which LLM is cheapest?" is the wrong question. The right one is "which model is cheapest for my workload, at the quality I need?" Two models with identical headline prices can cost you wildly different amounts depending on how you use them.

Sticker price isn't total price

Each provider publishes a per-million-token rate for input and output, and they're usually different from each other. But three things distort the real cost:

1. Input/output ratio. A model with cheap input but expensive output is great for extraction and terrible for long-form generation. Compare on your ratio, not a generic one.
2. Tokenizer efficiency. The same text tokenizes to different token counts across model families. A model that's 10% cheaper per token but uses 15% more tokens for your content is actually more expensive.
3. Reasoning / thinking tokens. Models that "think" before answering bill those tokens too. A higher effort setting can quietly multiply output cost.

Match the model to the job

• High-volume, simple tasks (classification, routing, tagging) → a small, fast, cheap model is usually plenty.
• Complex reasoning, agentic work → a flagship model earns its price by getting it right in fewer attempts.
• Long shared context → prompt caching can dominate the economics; factor it in.

Run the numbers on your traffic

Pick a representative request — real input and output token counts — and multiply by expected volume for each candidate model. That single comparison tells you more than any pricing-page skim.

The Token Cost Calculator makes this concrete: choose a model or enter any provider's exact per-million rates, plug in your typical token counts and request volume, and read the per-request and total cost. Do it for two or three candidates and the right choice for your workload usually becomes obvious.