The effective price of frontier models
Every provider prices in dollars per million tokens, which invites a comparison the unit can't support: a token is not a fixed amount of text. Each vendor trains its own tokenizer, so the same prompt becomes a different number of tokens depending on who's counting. Two models with identical sticker prices can bill you meaningfully different amounts for identical work.
The number that predicts your bill is effective price: list price × tokenizer divergence — how many tokens this model produces per token of a baseline (we use OpenAI's o200k, the most token-efficient of the majors). Try it live on your kind of content:
| Model | Tokens | vs o200k | List $/1M in | Effective $/1M in |
|---|---|---|---|---|
| GPT-5.6 Sol | … | … | $5 | — |
| Claude Fable 5 | … | … | $10 | — |
| Claude Opus 4.5 | … | … | $5 | — |
| Gemini 3 Pro | … | … | $2 | — |
| Grok 4.5 | … | … | $2 | — |
| Llama 3.3 70B | … | … | — | — |
| DeepSeek V3.1 | … | … | $0.28 | — |
Live — tokenized in your browser (rows marked ~est use a calibrated estimate until exact counts are available). Effective price = list price × tokens relative to o200k for this content.
Three patterns worth knowing. First, divergence depends on content: it's smallest on English prose and largest on code — camelCase identifiers, indentation, and JSX-style syntax are exactly what o200k compresses well and other vocabularies don't. If you're building a coding agent, the divergence on your repository is the one that matters, which is why the demo above (and the full comparison table) count your actual text instead of quoting an average.
Second, tokenizer upgrades are silent price changes. When a vendor ships a new model generation with a new tokenizer, the price sheet can stay identical while every request quietly costs more. Independent measurements of Anthropic's newest generation found its new tokenizer produces roughly 25–35% more tokens than its predecessor on the same content — verified against real billed requests (PlayCode's analysis). Same $/Mtok, ~30% higher bills. Comparing model generations in our table shows this directly, because every count comes from the model's actual tokenizer.
Third, divergence compounds. A tokenizer that produces 30% more tokens also makes every cache write, cache read, and context-window fill 30% more expensive — and long agent loops re-send their history every turn, multiplying the gap. Meanwhile output pricing, reasoning tokens, and model verbosity can move whole-task costs further than input tokenization ever will. $/Mtok is a fine opening line; it's just not the price.
The practical takeaway: paste a representative prompt — or a whole document — into the comparison table, set your monthly volume, and read the cost column. That's effective price, computed on your content, with each model's own tokenizer.