What $1 actually buys on an LLM API

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.

The conversion

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.

The number people get wrong: they read "714,000 tokens for a dollar" and assume that is 714,000 words of conversation. It is not. Token counts depend on the tokenizer, and the same text can produce noticeably different counts across model families. Measuring your own workload beats any conversion table, including this one.

What a dollar looks like as actual work

Converting to tasks is where it becomes intuitive. Three rough shapes, assuming a 2,000-token prompt and a 500-token answer per request:

WorkloadTokens per requestRequests 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.

Working out your own number

Four steps, and you can do them before spending anything:

  1. Count tokens, do not estimate. If you have logs, sum prompt_tokens and completion_tokens over a representative day. If you do not, send ten representative requests through any tokenizer and multiply.
  2. Separate the two columns. Most people only track "tokens" and then cannot explain their bill. Input and output are different line items with different rates.
  3. Multiply by the rate for the model you actually use, not the cheapest one in the catalogue. A cheap model you cannot use is not a saving.
  4. Add the retries. Real workloads retry. If your observed success rate is 97%, add roughly 3% to the total — and if it is much lower, that is the problem worth fixing first.
monthly ≈ (input_tokens/1e6 × in_rate + output_tokens/1e6 × out_rate) × 30 ÷ observed_success_rate

When one dollar is not enough

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.

Where Tapix AI fits

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

Last updated: 2026-11-21最后更新:2026-11-21