The output line on a token bill

Two requests can carry the same prompt and return equally useful answers while leaving different charges on the bill. The difference often sits in the output line: both the price of generated tokens and the number the model needs to finish the task. Comparing input rates alone misses that second half.

A useful side-by-side comparison keeps the workload fixed, separates input from output, and checks what the provider actually counts as billable generation.

§ Input price beside output price

Input tokens represent the material sent to a model: instructions, conversation history, documents, and tool results. Output tokens represent generated material. These categories can have different rates, so a single headline price is not enough to estimate a request.

• Reading-heavy work: A long document followed by a short classification makes input pricing especially important.
• Writing-heavy work: A brief instruction followed by a substantial draft makes output pricing especially important.

The basic calculation is input tokens multiplied by the input rate, plus output tokens multiplied by the output rate. When rates are quoted per million tokens, divide each token count by one million before multiplying. Cached input and other billing categories may need separate lines.

§ Two hypothetical rate cards, one workload

The following numbers are illustrative, not current Vana API prices or quotes for particular models. Assume standard input billing, with no cache discounts or additional charges.

• Option A: $1 per million input tokens and $5 per million output tokens.
• Option B: $2 per million input tokens and $3 per million output tokens.

For a request containing 10,000 input tokens and producing 500 output tokens, Option A costs $0.0125. Option B costs $0.0215. The lower input rate wins because the answer is short.

Keep those 10,000 input tokens, but increase the output to 10,000 tokens. Option A now costs $0.06, while Option B costs $0.05. The option with the higher input rate becomes cheaper overall.

At this fixed input length, the two options break even at 5,000 output tokens. The price gap is therefore a property of the workload, not just the rate card.

§ Equal token counts beside equal results

A rate comparison holds token counts constant. A practical evaluation holds the desired result constant. Those are different tests.

• Equal counts: Pricing the same number of input and output tokens isolates the rate difference.
• Equal results: Asking each model to produce an acceptable answer reveals differences in verbosity, retries, and tool use.

A cheaper output rate can lose its advantage if a model generates substantially more text or requires another attempt. Conversely, a higher-priced model may cost less per completed task if it reaches the required result with fewer billable tokens. Neither outcome should be assumed before testing.

Tokenization can also differ between models. The same English paragraph does not necessarily have the same token count everywhere, so compare recorded usage rather than estimating from word counts alone.

§ Visible text beside billable generation

The response on screen is not always the full accounting boundary. Depending on the model and billing rules, reasoning tokens may contribute to billable output even when they are not exposed as ordinary answer text. Inspect usage fields and the applicable pricing details rather than measuring only the final paragraph.

Likewise, an output limit is a ceiling, not a promise of exact spending. Its treatment of reasoning and visible text can vary. A concise response instruction is useful, but it does not replace checking actual usage.

§ A small comparison before a larger run

Vana API provides one OpenAI-compatible endpoint at https://api.vanaapi.com and also supports Anthropic Messages and Gemini native protocols. Its per-real-token billing, with no monthly fee, makes a small workload sample a practical starting point.

Create an account, run representative requests, and record input tokens, billable output tokens, total cost, and whether each result passed your acceptance criteria. Compare short answers separately from long drafts. The useful number is not simply the lowest output rate; it is the cost of an answer you can actually use.