GPT, Claude, and Gemini do not share a tokenizer or a rate card. Compare them on one prompt and one output assumption. Rates checked 12 Aug 2026.
Same prompt, every model
Paste once. TokenCalculator counts each family separately and applies that row’s published rates. That is more honest than ranking three list prices.
Published rates at a glance
USD per 1 million tokens from the curated catalog. Example math still belongs in the calculator, because token counts differ by family.
| Model | Provider | Input / 1M | Output / 1M |
|---|---|---|---|
| GPT-5 | OpenAI | $1.25/1M | $10.00/1M |
| Claude Sonnet 5 | Anthropic | $2.00/1M | $10.00/1M |
| Claude Opus 4.8 | Anthropic | $5.00/1M | $25.00/1M |
| Gemini 2.5 Pro | $1.25/1M | $10.00/1M | |
| Gemini 3.1 Pro | $2.00/1M | $12.00/1M | |
| GPT-4o mini | OpenAI | $0.150/1M | $0.600/1M |
| Claude Haiku 4.5 | Anthropic | $1.00/1M | $5.00/1M |
| Gemini 2.5 Flash | $0.300/1M | $2.50/1M |
What actually changes cost
- Output length: Claude and GPT flagship output rates are high. Long answers dominate.
- Tokenizer: the same paragraph can mint more tokens on one family than another.
- Cache: stable system prompts discount input when the catalog publishes a cache rate.
- Tier: Flash, Haiku, and Mini exist so most turns never touch Opus or GPT-5.
A fair comparison workflow
Save one golden prompt that looks like production: system rules, a user turn, and any tools you always send. Paste it once. Set the output size you expect, not a hopeful minimum. Record monthly volume. Then switch models without editing the text.
When each stack is usually cheaper
Gemini Flash and GPT Mini class rows often win high volume chat. Claude Haiku is the Anthropic volume tier. Frontier GPT-5, Claude Opus, and Gemini Pro are for hard turns where retries on a cheap model would cost more than a single expensive call.
Frequently asked questions
Who is cheapest overall?
There is no single winner. Use the cheapest rank for list price, then this compare table for your prompt. DeepSeek and host copies can undercut all three labs on open weights.
Should I trust in-browser counts?
OpenAI rows can be Exact. Claude and Gemini rows are Approx in the browser. For a launch budget, confirm Claude with Anthropic count_tokens and Gemini with Google countTokens.