AI Cost Simulator: what will your AI job cost?
Pick a job, such as a support chatbot, a coding assistant or invoice extraction. See what it costs per month on 386 priced models, which ones passed our benchmark, and how the bill grows as you scale. Change any number to match your own work.
- 386 models priced
- 41 benchmarked
- 13 use cases
- 24 benchmark prompts each
What common AI jobs cost, at a glance
Each row is a typical month for that job. The three columns show what it costs on the cheapest model that clears our quality bar, on one priced near the middle of the models that clear it, and on the flagship many people default to. Click a job to simulate it.
| AI job | Typical month | Budget pick | Middle pick | Flagship default | Flagship vs middle |
|---|---|---|---|---|---|
| Customer support chatbot A bot that answers customer questions from your help centre and hands hard cases to a person. | 3,000 conversations 9,000 in · 700 out each | $1.74 DeepSeek V4 Flash 0423 | $7.07 MiniMax M3 | $131 GPT-5.5 | 19x |
| Coding assistant for a team Chat, edits and explanations inside the editor for a small engineering team, billed by the token. | 4,400 requests 8,000 in · 1,200 out each | $4.12 DeepSeek V4.1 Flash | $35.29 Kimi K2.6 | $281 GPT-5.5 | 8x |
| Document summarisation Turn reports, contracts, articles or emails into a short summary or a set of key points. | 2,000 documents 9,000 in · 600 out each | $1.86 DeepSeek V4 Flash 0423 | $10.80 Mistral Large 3 2512 | $126 GPT-5.5 | 12x |
| Blog and content writing Drafts, rewrites and edits of blog posts, product pages and newsletters. | 120 articles 2,500 in · 3,000 out each | $0.083 Ministral 3 8B 2512 | $0.525 GLM 5.3 | $11.76 GPT-5.5 | 22x |
| Knowledge base Q&A (RAG) Staff or customers ask questions, the system fetches the right passages and the model answers from them. | 15K questions 4,500 in · 350 out each | $5.62 DeepSeek V4 Flash 0423 | $51.65 Kimi K2.6 | $404 GPT-5.5 | 7.8x |
| Invoice and form data extraction Pull fields such as supplier, date, totals and line items out of invoices, receipts and forms into JSON. | 20K documents 2,000 in · 400 out each | $4.65 DeepSeek V4 Flash 0423 | $49.85 Kimi K2.7 Code | $416 GPT-5.5 | 8.3x |
| Ticket and content classification Label a huge stream of tickets, reviews, emails or posts: topic, urgency, spam, language, sentiment. | 500K items 450 in · 12 out each | $11.30 DeepSeek V4.1 Flash | $91.13 Mistral Large 3 2512 | $1,001 GPT-5.5 | 11x |
| AI agent workflows An agent that plans, calls tools and loops until a task is done: research, triage, data clean-up, back-office jobs. | 1,500 runs 60K in · 7,000 out each | $7.44 DeepSeek V4 Flash 0423 | $71.67 Kimi K2.5 | $735 GPT-5.5 | 10x |
| Sales email personalisation Write a tailored opening line or a whole outreach email for each lead from a short profile. | 5,000 emails 1,100 in · 280 out each | $0.595 DeepSeek V4 Flash 0423 | $7.21 Kimi K2.7 Code | $59.60 GPT-5.5 | 8.3x |
| Meeting notes and summaries Turn call and meeting transcripts into minutes, action items and follow-up emails. | 400 meetings 14K in · 900 out each | $0.575 DeepSeek V4 Flash 0423 | $3.34 Mistral Large 3 2512 | $38.80 GPT-5.5 | 12x |
| Code review for pull requests An automated first-pass review on every pull request: bugs, risks and style. | 600 pull requests 16K in · 1,400 out each | $0.840 DeepSeek V4.1 Flash | $6.06 GLM 5.2 | $67.79 GPT-5.5 | 11x |
| Personal AI assistant for a founder Day-to-day chat for drafting, thinking things through, research and planning, for one person. | 1,500 messages 3,500 in · 700 out each | $0.478 DeepSeek V4 Flash 0423 | $5.68 Kimi K2.6 | $45.94 GPT-5.5 | 8.1x |
| Text-to-SQL data analyst Staff ask questions in plain English and the model writes the SQL against your schema. | 6,000 queries 3,500 in · 300 out each | $2.54 Qwen3.8 Flash | $14.84 Kimi K2.6 | $113 GPT-5.5 | 7.6x |
Estimates from the assumptions on each job's page, priced on the OpenRouter catalogue as of 2026-11-05. They are not quotes.
Simulate your own workload
Pick a job, set how much of it you do, and see what every model would cost. Defaults are typical, not yours. Change them.
A bot that answers customer questions from your help centre and hands hard cases to a person.
Prompt, history, documents
What the model writes
Advanced: caching, reasoning, retries, discounts
Share that can use cached-input prices where a model offers them
1 = none. 2 = hidden thinking doubles billed output
Extra calls on top of the base count
Batch jobs or committed spend. Check your provider; we assume none by default
Lowest benchmark score a model can have and still be picked. Higher is stricter.
Limit to certain providers
Untick any provider you cannot use, for data-residency, contract or policy reasons. Picks are recalculated.
3,000 conversations a month, three ways to pay
Every model here cleared a score of 9.0 out of 10 on our summarise, classify, rewrite benchmark tasks, ran without errors and fits a 32K-token window. 24 of 386 priced models qualify.
Cheapest model that clears the bar.
DeepSeek V4 Flash 0423
- Per conversation
- $0.0006
- Benchmark
- 9.56 / 10 · 9 prompts
- Context
- 1.05M tokens
Priced near the middle of the models that clear the bar; best score there.
MiniMax M3
- Per conversation
- $0.0024
- Benchmark
- 9.56 / 10 · 9 prompts
- Context
- 1.05M tokens
What many people reach for first.
GPT-5.5
- Per conversation
- $0.044
- Benchmark
- 9.56 / 10 · 9 prompts
- Context
- 1.05M tokens
Choosing GPT-5.5 over MiniMax M3 costs $124 more a month, $1,491 a year (19x the price).
Both scored 9.56 out of 10 on this kind of work. The extra money buys safety margin our benchmark cannot measure: it is 24 short prompts, and most passing models score near the top. If a wrong answer is costly for you, that margin can be worth it. Test on your own data before you decide.
If it grows
Pay-per-use cost scales in a straight line. Monthly cost at 1x, 3x, 10x and 100x today's volume.
| Model | 1x · 3,000 | 3x · 9,000 | 10x · 30K | 100x · 300K |
|---|---|---|---|---|
| DeepSeek V4 Flash 0423 | $1.74 | $5.23 | $17.43 | $174 |
| MiniMax M3 | $7.07 | $21.21 | $70.69 | $707 |
| GPT-5.5 | $131 | $394 | $1,314 | $13,136 |
What a budget buys
- DeepSeek V4 Flash 0423172Kconversations a month, about 5,737 a day
- MiniMax M342Kconversations a month, about 1,414 a day
- GPT-5.52,283conversations a month, about 76 a day
Every model, priced for your workload
41 models, 15 hidden because their window is smaller than 32K tokens.
| Model | A month | Per conversation | Benchmark | Context |
|---|---|---|---|---|
| DeepSeek V4 Flash 0423 DeepSeekBudget pick | $1.74 | $0.0006 | 9.56 (9) | 1.05M |
| DeepSeek V4.1 Flash DeepSeekbelow the bar | $1.92 | $0.0006 | 7.89 (9) | 1.05M |
| Ministral 3 8B 2512 Mistral AI | $2.29 | $0.0008 | 9.44 (9) | 262K |
| Llama 4 Scout Meta | $2.73 | $0.0009 | 9.44 (9) | 1.31M |
| Qwen3.8 Flash Alibaba | $3.01 | $0.0010 | 9.33 (9) | 1M |
| Mistral Small 4 Mistral AIa benchmark run failed | $3.28 | $0.0011 | 9.63 (8) | 262K |
| GLM 5.3 Flash Z Ai | $3.31 | $0.0011 | 9.44 (9) | 1.05M |
| GLM 5.3 Z Ai | $5.05 | $0.0017 | 9.67 (9) | 1.05M |
| GPT-5.4 Nano OpenAI | $5.36 | $0.0018 | 9.56 (9) | 400K |
| MiniMax M2.5 Minimax | $5.67 | $0.0019 | 9.44 (9) | 205K |
| Qwen3.8 27B Alibababelow the bar | $6.47 | $0.0022 | 7.78 (9) | 1.02M |
| Muse Glimmer 30B Metabelow the bar | $6.73 | $0.0022 | 8.89 (9) | 131K |
Benchmark = mean judge score out of 10 on the summarise, classify, rewrite prompts of the Model Playground, with the number of prompts in brackets. Models marked n/a have no benchmark run, so we do not guess their quality. How the benchmark works.
How the estimate works
The cost. For each model we multiply the tokens one job sends by the model's input price, add the tokens it gets back at the output price, then multiply by how many jobs you run. Repeated input uses a model's cached-input price when it has one. Reasoning, retries and any discount you enter scale the result.
The prices. They come from the OpenRouter model catalogue we sync into our own data. Free previews, duplicate listings and retired models are left out. We never guess a batch discount from a listing: you enter the discount you actually get.
The quality bar. Each job maps to the kinds of tasks in our Model Playground, a set of 24 prompts across eight task types that we ran on 41 models and scored with claude-haiku-4.5. A model is a pick only if it ran without errors, fits the job's context window and reaches the bar for that job.
What the benchmark can and cannot tell you. There are three prompts per task type, so one hard prompt moves a score a lot. Most models that pass land close to the top of the scale. It is good at separating models that fail from models that pass, and weak at ranking the ones that pass. That is why we show a ladder of prices and not a single winner.
What is not included. Embeddings, vector databases, speech-to-text, image inputs, warehouse compute, engineering time and payment fees. Long multi-step runs fail more often than any single step suggests. Treat every number as an estimate to test, not a quote.
Before you switch. Run your own 100 to 300 real examples on the model you are considering and count the wrong answers. A model that is 20 times cheaper but wrong one time in ten may cost more once you pay for the fixes.
Questions people ask
How much does it cost to run AI for my business?
It comes down to three numbers: how many jobs you run, how long each one is, and which model does the work. At the default workloads on this page, the cheapest model that clears our benchmark costs between $0.083 and $11.30 a month across 13 common jobs. A flagship model on the same jobs costs between $11.76 and $1,001. Change the volume and token counts to see your own bill.
Where do the prices come from?
From the OpenRouter model catalogue, which we sync into our own data (latest price check 2026-11-05). Vendors can change prices at any time, so confirm on the provider's own pricing page before you commit. The page shows the date the prices were last checked.
How does the simulator choose which models to show?
It prices every metered API model in our catalogue on your workload, then looks at the ones we have benchmarked in the Model Playground (24 prompts across eight task types, scored by claude-haiku-4.5). A model counts only if it ran without errors, fits the job's context window and reaches the quality bar for that job. The cheapest of those is the budget pick, one priced near the middle of that group is the middle pick, and the flagship is what many people default to.
Why do cheap models score as well as expensive ones?
Most models that pass our benchmark score close to the top of the scale, so it separates models that fail from models that pass better than it ranks the passing ones. We show the price at three points on that ladder instead of naming one winner. A flagship may still earn its price on long, messy or high-stakes work the benchmark does not cover.
Does it include ChatGPT Plus, Claude Pro and other subscriptions?
For jobs where a person uses the output directly, such as writing, coding and a personal assistant, the simulator also lists per-seat subscription plans and the volume at which paying per use costs the same. Automated jobs such as classification, extraction and support bots cannot run on a seat plan, so no subscription is shown.
More ways to check what you pay
- AI Value FrontierWhich models give you the most measured quality per dollar, and which ones you can drop.
- Decision Model Cost LabLabelling millions of items? See when a decision model beats asking a chat model.
- AI Cost CalculatorAlready know your model and token counts? Get an exact monthly figure and cheaper swaps.
- Model PlaygroundThe benchmark behind the scores: same prompt, every model in a family, priced per 1,000 runs.
- Latest AI releasesNew models as vendors announce them. Price a new one here once it is listed.
- Model retirementsModels being switched off, and what moving off them will cost.