Input vs Output Token Pricing — Why Output Costs 3–5× More
Output tokens cost significantly more than input tokens on every LLM API. This isn't arbitrary — it reflects the fundamental difference in compute required. Understanding this changes how you prompt.
The pricing gap
Claude Opus 5: $15/M input, $75/M output (5× more). GPT-5.4: $2.50/M input, $10/M output (4× more). Gemini 2.5 Pro: $1.25/M input, $10/M output (8× more). The asymmetry is consistent across all providers and has been stable for 3+ years.
Why output costs more: the compute reason
Input tokens run through the model's forward pass once in parallel (fast). Output tokens are generated auto-regressively — one token at a time, each requiring a full forward pass. For a 500-token output, that's 500 sequential forward passes. This is fundamentally more compute-intensive than a single parallel input forward pass.
Practical implications for prompt design
1. Ask for concise answers: "Answer in under 50 words" saves 80% on output cost vs open-ended. 2. Use JSON for structured extraction — less verbose than prose. 3. Avoid "Sure, I'd be happy to help! Here's what I found..." — instruct the model to skip preambles. 4. Multiple questions in one call (share the output cost) is cheaper than separate calls.
When long output is worth paying for
Code generation, document drafting, and detailed analysis legitimately require long outputs. In these cases, minimize input costs instead (shorter prompts, caching) rather than trying to compress the output. Don't optimize the wrong end.
Related calculators
Related guides
Frequently Asked Questions
How do I estimate my output token costs?
Most models generate roughly the amount they're asked for. A 200-word answer ≈ 267 tokens. If you're getting 500-token responses on average and making 100K calls/month, that's 50M output tokens. At $10/M (GPT-5.4), that's $500/month from output alone.
Guide updated 2026-08-20. Pricing data verified weekly. See methodology · sources.