Workload costs
Nobody knows what two million tokens a day means. So pick the job and how often you run it — the bill comes out per model, as a range, with every assumption visible below. Ranges, not single numbers: the inputs are ranges too, and a decimal here would be a guess in a suit.
| Model | Class | $/day | $/month | Reasoning | Coding | Agentic |
|---|
Straight per-token arithmetic on list prices: no caching, no batching discount, no volume deal. Thinking tokens bill as output, so treat reasoning models as a floor. Weight class is kept because class also tells you something about quality. Capability & cost data: Artificial Analysis
THE ASSUMPTIONS
Every number above rests on these. They are rules of thumb, published so you can disagree with them:
- 1 page of A4 text ~ 500-800 tokens
- 1 business email ~ 200-400 tokens; with a quoted thread ~ 800-1,500
- 1 scanned invoice or receipt (OCR text) ~ 400-900 tokens
- 1 support conversation turn ~ 150-300 tokens
- Retrieval-augmented answers add the retrieved chunks: 3-6 chunks ~ 1,500-3,000 tokens
| Task archetype | Tokens in | Tokens out | Basis |
|---|---|---|---|
| Classify & routeRead one short item and pick a label or destination — inbox triage, ticket routing, lead scoring. | 800–1,800 | 120–300 | one email or ticket including a short quoted thread; output is a label plus a one-line reason |
| Extract fields from a documentPull structured data out of an invoice, receipt, order or contract page. | 600–1,600 | 150–400 | one OCR'd page; output is a JSON record of 8-20 fields |
| Draft a replyWrite a customer-ready answer grounded in the incoming message and a short context block. | 1,200–3,000 | 250–600 | incoming message plus tone/policy context; output is a full email |
| Answer from a knowledge baseRetrieval-augmented answer over your own documents — SOP lookup, policy question, product detail. | 2,500–6,000 | 250–700 | question plus 3-6 retrieved chunks; output is a cited answer |
| Summarise something longCondense a meeting transcript, report or thread into a briefing. | 6,000–20,000 | 400–1,200 | a 30-60 minute transcript or a 10-25 page document |
| Write a piece of contentProduce a post, newsletter section or product description from a brief and source material. | 2,000–8,000 | 800–2,500 | brief plus reference material; output is publishable copy |
| Multi-step agent runA goal the model pursues over several tool calls — each step re-reads the growing context. | 15,000–60,000 | 1,500–6,000 | 6-15 tool-calling steps with accumulated context; the dominant cost is re-reading, not writing |
| Scheduled checkA small recurring evaluation — did anything change, does this look wrong, is this worth alerting. | 400–1,200 | 80–250 | one status payload or diff; output is a verdict plus a short reason |
Blueprints that run no model at all (The chargeback evidence pack, The house AI tool, The gift voucher register, The subcontractor payment gate, The hiring inbox that answers everyone, The allergen matrix that stays right, The goods-in check, The warranty claim register…) cost hosting, not tokens — they are not in the picker, because a token bill for them would be fiction.