Per-million-token prices are hard to feel. A dollar is easier. Below is a dollar converted into tokens at our current prices, measured against the live catalogue on 21 November 2026 — and a way to work out what your own workload costs before you top up anything.
A price of $X per million tokens means one dollar buys 1 ÷ X × 1,000,000 tokens. That is the whole formula. Input and output are priced separately, so you get two different numbers per model:
| Model | $1 buys (input) | $1 buys (output) |
|---|---|---|
| claude-sonnet-5 | ≈ 714,000 tokens | ≈ 143,000 tokens |
| gpt-5.5 | ≈ 286,000 tokens | ≈ 48,000 tokens |
| gemini-3-flash | ≈ 2,857,000 tokens | ≈ 476,000 tokens |
Those numbers move with the price list, not with this page — the pricing page and the calculator are the live versions. What does not move is the structure: on every model, output costs several times input, so the output column is the one that decides your bill.
Converting to tasks is where it becomes intuitive. Three rough shapes, assuming a 2,000-token prompt and a 500-token answer per request:
| Workload | Tokens per request | Requests per $1 (claude-sonnet-5) |
|---|---|---|
| Short Q&A with modest context | ≈ 2,500 | ≈ 250 |
| Code review of one file | ≈ 12,000 | ≈ 50 |
| Long-document summary (30k context) | ≈ 31,000 | ≈ 20 |
Notice how fast the request count falls as context grows. That is the single most useful thing to internalise about API pricing: **cost scales with context, not with the number of questions.** A chat that keeps its history gets more expensive per turn, not cheaper.
Four steps, and you can do them before spending anything:
prompt_tokens and completion_tokens over a representative day. If you do not, send ten representative requests through any tokenizer and multiply.monthly ≈ (input_tokens/1e6 × in_rate + output_tokens/1e6 × out_rate) × 30 ÷ observed_success_rate
A dollar is enough to answer "does this work at all". It is not enough to answer "does this work at my volume". The honest thresholds:
There is also a reason we cap the free starting credit at $1 rather than handing out more: a larger free tier attracts volume that never converts, and the cost of that gets priced into everyone else's rates. A dollar is enough to test and small enough to give away honestly.
The numbers above are our prices, and the same arithmetic applies to any provider — the three checks exist so you can confirm the price you were charged without trusting a table. If your own measurement disagrees with a rate on our pricing page, that is a bug on our side and we want the request ID.
Start with a dollar, measure, then decide. That order is the whole point.
Run the numbers in the calculator · See live prices vs official list · Free tiers compared
Keep reading: Free tiers compared · LLM API price comparison · Cheap LLM API options