How many tokens does one dollar buy?
At a 4:1 input-to-output ratio — roughly what agent and chat traffic runs at — one dollar buys about 3.5M tokens on MiniMax-M2.7, 1.7M on Kimi K2.6, and about 0.3M on Claude Sonnet 5. The spread between an open-weight model and a frontier one is roughly an order of magnitude, which is why the same budget behaves so differently depending on what you point it at.
The numbers
| Model | $/1M in | $/1M out | Blended 4:1 | Tokens per $1 |
|---|---|---|---|---|
| MiniMax-M2.7 | $0.18 | $0.72 | $0.288 | 3.5M |
| Kimi-K2.6 | $0.35 | $1.46 | $0.572 | 1.7M |
| GLM-5.2-FP8 (paid plans) | $0.38 | $1.19 | $0.542 | 1.8M |
| Claude Sonnet 5 | $2.00 | $10.00 | $3.600 | 0.3M |
| GPT-5.4 | $2.50 | $15.00 | $5.000 | 0.2M |
| Claude Opus 5 | $5.00 | $25.00 | $9.000 | 0.1M |
Why 4:1
Per-token prices are quoted separately for input and output, and output costs several times more, so a single "price" only means something once you fix a ratio. Chat and agent traffic sends far more than it generates — long system prompts, tool results and conversation history on every turn — and 4:1 is a reasonable middle. If your workload writes long documents from short prompts, your effective price is higher than this table; recompute with your own ratio.
blended = (4 * input_price + 1 * output_price) / 5
tokens_per_dollar = 1_000_000 / blendedWhat this misses
- Reasoning tokens. Models that think before answering bill those as output, and they do not appear in your prompt or your reply. A model with a lower sticker price can cost more per finished task.
- Retries and failed runs. An agent that loops twice costs twice, at any price.
- Cached input. Some providers discount repeated prefixes heavily; if your prompts share a long prefix, that can matter more than the headline rate.
- Subscriptions. A plan is a fixed monthly amount against an allowance, which changes the arithmetic — see the pricing page.
Prices for Claude and GPT come from the vendors’ own pricing pages; ours are what we bill. Sources and the dated table are on the pricing page.
FAQ
▸How many words is a token?
Roughly 0.75 words in English, so 1M tokens is on the order of 750,000 words. Code and non-English text tokenise less efficiently — often noticeably so.
▸Which model gives the most tokens per dollar?
Of the models we serve, GLM-5.2 at $0.38/$1.19 — but it is on paid plans only. Of the models on every plan, MiniMax-M2.7.
▸Is a subscription cheaper than paying per token?
It is if you use it. A plan holds a money allowance that buys tokens at our per-token rates, sized so a plan is worth several times its price at those rates — the value comes from actually spending the allowance. Occasional use is cheaper pay-as-you-go.