Cost Reference
Every AI cost calculator asks for a token count. Nobody knows their token count — they know they have a 100-page PDF and a question about the bill. So each job below carries a defensible token estimate, priced across every model in the directory, with the working shown.
Pricing verified:
| Task | Cheapest | Best value | Premium | Tokens in / out |
|---|---|---|---|---|
| Documents | ||||
| Summarize a 100-page PDF | $0.0036 GPT-5 Nano | $0.0036 GPT-5 Nano | $0.690 GPT-6 Astra | 65,000 / 800 |
| Summarize a 400-page book | $0.0086 GPT-5 Nano | $0.0086 GPT-5 Nano | $1.68 GPT-6 Astra | 160,000 / 1,500 |
| Translate a 50-page document | $0.0068 Ministral 3 3B 2512 | $0.045 DeepSeek V4-Pro | $0.425 Claude Sonnet 5 | 32,500 / 36,000 |
| Writing | ||||
| Write a 1,500-word blog post | $0.0002 Gemma 2 9B | $0.0023 DeepSeek V4-Pro | $0.022 Claude Sonnet 5 | 1,200 / 2,000 |
| Write a tailored cover letter | $0.0001 Gemma 2 9B | $0.0011 GPT-5.6 Luna | $0.0099 Claude Sonnet 5 | 2,200 / 550 |
| Write 1,000 product descriptions | $0.0000 Gemma 2 9B | $0.0003 DeepSeek V4-Pro | $0.0027 Claude Sonnet 5 | 250 / 220 |
| Engineering | ||||
| Review a 500-line pull request | $0.0005 Llama 3.2 1B Instruct | $0.0011 Mistral Small 3.2 24B | $0.082 Claude Opus 4.8 | 12,000 / 900 |
| Support | ||||
| Summarize a 1-hour meeting | $0.0005 Llama 3.2 1B Instruct | $0.0032 GPT-5.6 Luna | $0.031 Claude Sonnet 5 | 12,000 / 700 |
| Answer 5,000 support tickets | $0.0001 Gemma 2 9B | $0.0011 GPT-5.6 Luna | $0.011 Claude Sonnet 5 | 3,500 / 350 |
| Run a chatbot for 10,000 chats | $0.0003 Gemma 2 9B | $0.0023 GPT-5.6 Luna | $0.021 Claude Sonnet 5 | 6,000 / 900 |
| Data | ||||
| Extract data from 1,000 invoices | $0.0001 Gemma 2 9B | $0.0003 DeepSeek V4-Flash | $0.027 Claude Fable 5 | 1,200 / 300 |
| Analyze a 10,000-row spreadsheet | $0.0050 GPT-5 Nano | $0.013 DeepSeek V4-Flash | $0.960 Claude Fable 5 | 90,000 / 1,200 |
Costs are per single run. “Best value” is the cheapest model that still scores 78+ on the dimension that matters for that task; “premium” is the highest-scoring model regardless of price. Open a task for monthly totals at realistic volume.
Output tokens usually cost three to five times input tokens, so the shape of the job changes the ranking. Summarizing is input-heavy: you send 65,000 tokens and get 800 back, and the input price decides the bill. Translating is output-heavy: you pay full output rates on the whole document. A model that looks cheap on one can be the expensive choice on the other.
1,000 tokens ≈ 750 English words. A page of dense prose ≈ 650 tokens; a page of a book ≈ 400; a minute of speech ≈ 200; a spreadsheet row ≈ 9. Estimates assume you send the whole document rather than a retrieved slice, because that is what people do first. Every task page shows its own working.
Far less per task than most people expect, and far more at volume than most people plan for. summarize a 100-page pdf costs $0.0036–$0.690; summarize a 400-page book costs $0.0086–$1.68; translate a 50-page document costs $0.0068–$0.425. The per-run figure is almost never the problem — the monthly figure at real volume is.
API pricing spans roughly two orders of magnitude, from about $0.07 per million input tokens at the budget end to $15 or more at the frontier, and output tokens usually cost three to five times input. So a task's cost depends on both which model you pick and whether the job is input-heavy (summarizing) or output-heavy (translating, drafting).
From stated conventions rather than guesses: 1,000 tokens is roughly 750 English words, a page of dense prose is about 650 tokens, a page of a book about 400, a minute of speech about 200, and a spreadsheet row about 9. Each task page shows the full working so you can check it against your own documents.
No, deliberately. Batch APIs typically cut costs by half for work that can wait, and prompt caching can cut the input side by up to 90% when the same context is reused across calls. Leaving both out means every figure here is a ceiling rather than a best case.
Every number is computed from our model pricing data, which is checked daily against each provider's official pricing page. When a provider cuts a price, these figures fall with it on the next refresh.
Newsletter
Pricing changes, new model releases, and updated recommendations — delivered when it matters.
No spam. Useful updates only. Affiliate disclosures always clearly labeled.