Up to 70% off kits30% off services
AI Token Cost Calculator
Price an AI agent or LLM workload on 9 Claude, GPT and Gemini models at once. Enter the tokens one run uses across all its model calls and how many runs you expect a month. Prices are the providers' standard rates, checked 22 Nov 2026.
Your workload
Starting points to edit, not benchmarks. Your real numbers come from a test run.
Rule of thumb: 1,000 tokens is about 750 English words, so a 300-word email is about 400 tokens. Instructions and tool definitions are sent again on every call.
Cheapest of 9 models
$5.60 / month
Claude Haiku 5.5 · $0.0028 per run
The priciest model here, Claude Fable 5.2, costs $448 a month for the same work: about 80 times more.
| Model | Price per 1M (in / out) | Per run | Per month | Per year |
|---|---|---|---|---|
| Claude Haiku 5.5Anthropic · SmallPrompts up to 100,000 tokens | $0.10 / $0.50 | $0.0028 | $5.60 | $67.20 |
| GPT-6 LunaOpenAI · Small | $0.10 / $0.50 | $0.0028 | $5.60 | $67.20 |
| Gemini 3.5 Flash-LiteGoogle · Small | $0.30 / $2.50 | $0.01 | $20.80 | $250 |
| Gemini 3.8 FlashGoogle · MidRises to $1.50 in / $7.50 out on 1 Jan 2027 | $0.75 / $3.75 | $0.02 | $42.00 | $504 |
| Claude Sonnet 5.5Anthropic · Mid | $2 / $10 | $0.06 | $112 | $1,344 |
| GPT-6.2 SolOpenAI · Mid | $2 / $10 | $0.06 | $112 | $1,344 |
| Claude Opus 5.5Anthropic · Large | $4 / $20 | $0.11 | $224 | $2,688 |
| GPT-6.2 AstraOpenAI · Large | $8 / $40 | $0.22 | $448 | $5,376 |
| Claude Fable 5.2Anthropic · Large | $8 / $40 | $0.22 | $448 | $5,376 |
Want real numbers instead of estimates?30% OFF
Our $349$499 opportunity assessment runs your task on 50 real past examples with two models and reports the measured accuracy and cost per run, credited toward the build.
How to estimate tokens for an AI agent
A chat prompt is one call. An agent is several: it reads the item, looks something up, decides and writes. Each call resends the instructions, the tool definitions and what came before, so the input adds up faster than people expect.
- Count the calls in one run. Our worked example has 4: read the ticket, look up the customer, tag and route, draft the reply.
- Add up the input across calls. Instructions and tool definitions on every call, plus the item and the context that grows as the run goes on. The example uses 18,000 in total.
- Estimate the output. Tags and decisions are short; drafted replies are longer. The example writes 2,000.
- Cut what you don't need. Send only the fields the task uses, cache the instructions that repeat, and use a plain rule instead of a model wherever no reading or judgment is needed.
- Weigh it against the time saved. Model usage is usually the small bill. The AI agent ROI calculator puts it next to hours saved and the build cost, and our AI agent cost guide covers hosting and upkeep.
Frequently asked questions
Tokens, agents, where the prices come from and picking a model.
A token is a chunk of text the model reads or writes, usually part of a word. As a rough rule for English, 1,000 tokens is about 750 words. Providers charge separately for input tokens (everything you send: instructions, tool definitions, the email or document) and output tokens (what the model writes back). Output always costs more.
An agent usually makes several model calls per run, for example read the ticket, look up the customer, decide, then draft. Each call resends the instructions and earlier context, so input grows as the run goes on. Enter the total tokens across all calls in one run.
From each provider's own pricing page (Anthropic, OpenAI and Google), standard short-context rates, checked 22 Nov 2026. Longer prompts, cached input and batch jobs are priced differently. Gemini 3.8 Flash has a listed price rise on 1 January 2027. Always check the provider's page before you commit a budget.
Anthropic's Batch API charges 50% less for input and output when a job can wait for results. The toggle applies that discount to the Claude models only. OpenAI and Google also list batch prices; check their pages for current rates.
Not without testing. If a small model gets more answers wrong and a person has to fix them, the time spent costs more than the tokens saved. Run two or three models on the same 50 real examples and compare accuracy and cost together.