AI Prompt Cost Calculator: Estimate Token Costs Across Models (2026)

Two numbers drive every estimate: how many tokens you send (input) and how many the model returns (output). Each has its own price, quoted per 1,000,000 tokens.

The formula:

``` cost = (input_tokens / 1,000,000) * input_price_per_M + (output_tokens / 1,000,000) * output_price_per_M ```

To estimate token counts before you have an exact tokenizer count, use the rule of thumb that 1 token is approximately 4 characters or about 0.75 words in English (a rough estimate per OpenAI and Anthropic documentation). So a 500-word prompt is roughly 500 ÷ 0.75 ≈ 667 input tokens. This is an approximation; whitespace, punctuation, code, and non-English text shift the ratio.

Output dominates many bills because output prices are typically several times the input price, and long generations add up fast. If a task can return a short structured answer instead of prose, that alone cuts cost.

Say you send a 750-word prompt and get back a 750-word answer. At ~0.75 words per token, that's about 1,000 input tokens and 1,000 output tokens — 0.001 M each.

On gpt-5.4 ($2.50 input / $15.00 output per 1M): input = 0.001 × $2.50 = $0.0025; output = 0.001 × $15.00 = $0.015; total ≈ $0.0175 per call.

On Claude Sonnet 4.6 ($3.00 / $15.00): input = 0.001 × $3.00 = $0.003; output = 0.001 × $15.00 = $0.015; total ≈ $0.018 per call.

On Gemini 2.5 Flash ($0.30 / $2.50): input = 0.001 × $0.30 = $0.0003; output = 0.001 × $2.50 = $0.0025; total ≈ $0.0028 per call.

Same workload, roughly 6x cheaper on Gemini 2.5 Flash than on the mid-tier OpenAI or Anthropic models — which is the whole point of matching model tier to task difficulty.

Now run that same 1,000-in / 1,000-out call 100,000 times — say a batch classification or summarization job. Multiply the per-call totals above by 100,000:

gpt-5.4: $0.0175 × 100,000 ≈ $1,750. Claude Sonnet 4.6: $0.018 × 100,000 ≈ $1,800. Gemini 2.5 Flash: $0.0028 × 100,000 ≈ $280.

At this scale, discounts matter. If the job is not latency-sensitive, Anthropic's Batch API gives 50% off both input and output, halving the Sonnet figure to roughly $900. Prompt caching helps when a large, identical prefix (system prompt, instructions, reference doc) repeats across calls. See the methods below.

Two mechanisms can sharply reduce cost when your workload fits them.

Prompt caching (Anthropic): when many calls share the same large prefix, you cache it once and pay a reduced rate on the cached portion of later calls. Per Anthropic's pricing, a cache read (a hit) costs 0.1x the base input price — that is, 10% of the normal input rate for the cached tokens. Writing to the cache costs more than base input (1.25x for a 5-minute cache, 2x for a 1-hour cache), so caching pays off when the same prefix is reused enough times to amortize that write. Example: Claude Opus 4.8 input is $5/M, and its cache read is $0.50/M — a 90% saving on the repeated portion.

Batch API (Anthropic): 50% off both input and output for asynchronous, non-time-sensitive jobs. This stacks cleanly onto large offline workloads like the 100,000-call example above.

Match the discount to the shape of the work: caching for a big repeated prefix with varying tail, batch for large jobs you can wait on. Confirm current rates and mechanics on the linked pricing pages, since these terms change.

For planning, character or word counts get you close. Take your prompt's character count and divide by 4, or its word count and divide by 0.75, to estimate input tokens. Estimate output tokens from the length of answer you expect.

Then plug both into the formula and the price for your chosen model. Treat the result as an order-of-magnitude estimate, not a billing guarantee — the exact figure depends on the model's tokenizer and your actual output length. To draft tighter prompts (fewer tokens, clearer instructions) for whichever model you choose, our ChatGPT prompt generator and code prompt builder help keep inputs lean.