What Does AI Actually Cost Per Task? Real-World Examples
Pricing pages show you cost per million tokens. But nobody thinks in millions of tokens. You think in tasks: "How much does it cost to summarize this document?" or "What's the price of generating a product description?" This guide answers those questions with real numbers, real token counts, and real comparisons across every major provider.
We measured actual token usage for 12 common AI tasks, then calculated the exact cost across 20+ models. No estimates, no hand-waving — just the math.
💡 Key Takeaway: Most individual AI tasks cost fractions of a cent. The costs add up at scale — a customer support bot handling 10,000 conversations per day can range from $6/day to $500/day depending on which model you choose.
How We Calculated These Costs
Every cost in this article uses the same formula:
Cost = (Input Tokens × Input Price) + (Output Tokens × Output Price)
Token counts come from real-world usage — we ran each task through multiple models and averaged the token consumption. Input prices include the prompt, system instructions, and any context. Output prices cover the model's response.
All pricing data comes from official API rates as of March 2026, pulled directly from our calculator. No cached pricing, no promotional rates — just standard API costs.
Task 1: Summarizing an Email (Short Text)
Typical tokens: ~200 input, ~100 output
This is the simplest AI task — take a short email and produce a 2-3 sentence summary. It's also one of the cheapest.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| GPT-5 nano | $0.00001 | $0.00004 | $0.00005 |
| Gemini 2.0 Flash-Lite | $0.000015 | $0.00003 | $0.000045 |
| Mistral Small 3.2 | $0.000012 | $0.000018 | $0.00003 |
| GPT-5 mini | $0.00005 | $0.0002 | $0.00025 |
| Claude Haiku 4.5 | $0.0002 | $0.0005 | $0.0007 |
| GPT-5.2 | $0.00035 | $0.0014 | $0.00175 |
| Claude Opus 4.6 | $0.001 | $0.0025 | $0.0035 |
At $0.00003 per email with Mistral Small 3.2, you could summarize 33,000 emails for a single dollar. Even with Claude Opus 4.6 — the most expensive option — you'd still get 285 summaries per dollar.
📊 Quick Math: A company processing 1,000 emails/day would spend $0.03/day with Mistral Small or $3.50/day with Claude Opus. That's $11/year vs $1,277/year — a 116x cost difference for the same task.
Task 2: Customer Support Response
Typical tokens: ~500 input (customer message + context), ~300 output
Customer support is the highest-volume AI use case for most businesses. The input includes the customer's message, conversation history, and product knowledge base context. The output is a helpful response.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| GPT-5 nano | $0.000025 | $0.00012 | $0.000145 |
| Gemini 2.5 Flash | $0.000075 | $0.00018 | $0.000255 |
| DeepSeek V3.2 | $0.00014 | $0.000126 | $0.000266 |
| Mistral Small 3.2 | $0.00003 | $0.000054 | $0.000084 |
| GPT-5 mini | $0.000125 | $0.0006 | $0.000725 |
| Claude Sonnet 4.6 | $0.0015 | $0.0045 | $0.006 |
| GPT-5.2 | $0.000875 | $0.0042 | $0.005075 |
[stat] $0.000084 Cost of a single customer support response with Mistral Small 3.2
For a support desk handling 10,000 tickets per day, here's the monthly bill:
| Model | Daily Cost | Monthly Cost |
|---|---|---|
| Mistral Small 3.2 | $0.84 | $25 |
| GPT-5 nano | $1.45 | $44 |
| DeepSeek V3.2 | $2.66 | $80 |
| GPT-5 mini | $7.25 | $218 |
| GPT-5.2 | $50.75 | $1,523 |
| Claude Sonnet 4.6 | $60.00 | $1,800 |
The difference between the cheapest and most expensive option is 72x. For most support tickets — password resets, shipping status, return policies — you don't need a frontier model. Route simple tickets to a budget model and escalate complex ones to a smarter model.
💡 Key Takeaway: Implement tiered routing for customer support. Use Mistral Small or GPT-5 nano for straightforward queries (80% of tickets) and GPT-5.2 or Claude Sonnet for complex issues (20%). This alone can cut your AI support bill by 60-70%.
Task 3: Blog Post Generation (~1,500 words)
Typical tokens: ~300 input (topic + instructions), ~2,000 output
Generating long-form content is output-heavy, which means the output price matters far more than the input price. A 1,500-word blog post consumes roughly 2,000 output tokens.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.000018 | $0.00036 | $0.000378 |
| GPT-5 nano | $0.000015 | $0.0008 | $0.000815 |
| DeepSeek V3.2 | $0.000084 | $0.00084 | $0.000924 |
| Gemini 2.5 Flash | $0.000045 | $0.0012 | $0.001245 |
| GPT-5 mini | $0.000075 | $0.004 | $0.004075 |
| Gemini 2.5 Pro | $0.000375 | $0.02 | $0.020375 |
| GPT-5.2 | $0.000525 | $0.028 | $0.028525 |
| Claude Sonnet 4.6 | $0.0009 | $0.03 | $0.0309 |
| Claude Opus 4.6 | $0.0015 | $0.05 | $0.0515 |
Even at the premium end, generating a blog post costs 5 cents. If you publish 30 posts per month with Claude Opus, that's $1.55/month in API costs. Content generation is one area where quality usually matters more than cost — the price difference is negligible at low volume.
When volume matters: Content farms generating 1,000+ posts per day should optimize aggressively. At 1,000 posts/day, the spread is $11/month (Mistral Small) vs $1,545/month (Claude Opus).
Task 4: Code Generation (Function/Module)
Typical tokens: ~400 input (spec + context), ~800 output
Generating a utility function or small module based on a specification. This is a core developer workflow — describe what you want, get working code back.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.000024 | $0.000144 | $0.000168 |
| DeepSeek V3.2 | $0.000112 | $0.000336 | $0.000448 |
| GPT-5 mini | $0.0001 | $0.0016 | $0.0017 |
| Gemini 2.5 Flash | $0.00006 | $0.00048 | $0.00054 |
| Codestral | $0.00012 | $0.00072 | $0.00084 |
| GPT-5.2 | $0.0007 | $0.0112 | $0.0119 |
| Claude Sonnet 4.6 | $0.0012 | $0.012 | $0.0132 |
| Claude Opus 4.6 | $0.002 | $0.02 | $0.022 |
A developer generating 50 code snippets per day would spend:
- Mistral Small 3.2: $0.25/month
- GPT-5 mini: $2.55/month
- Claude Sonnet 4.6: $19.80/month
- Claude Opus 4.6: $33.00/month
⚠️ Warning: Code generation costs multiply fast when you add context. Sending an entire codebase as context (10,000+ tokens) can increase the input cost by 25x. Use targeted context — only send relevant files, not everything.
Task 5: Document Analysis (Long Document)
Typical tokens: ~8,000 input (document + question), ~500 output
Analyzing a contract, research paper, or financial report. This is input-heavy — the document itself dominates the token count.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Gemini 2.0 Flash-Lite | $0.0006 | $0.00015 | $0.00075 |
| GPT-5 nano | $0.0004 | $0.0002 | $0.0006 |
| Mistral Small 3.2 | $0.00048 | $0.00009 | $0.00057 |
| Gemini 2.5 Flash | $0.0012 | $0.0003 | $0.0015 |
| DeepSeek V3.2 | $0.00224 | $0.00021 | $0.00245 |
| GPT-5 mini | $0.002 | $0.001 | $0.003 |
| GPT-5.2 | $0.014 | $0.007 | $0.021 |
| Claude Sonnet 4.6 | $0.024 | $0.0075 | $0.0315 |
| Gemini 3 Pro | $0.016 | $0.006 | $0.022 |
| Claude Opus 4.6 | $0.04 | $0.0125 | $0.0525 |
📊 Quick Math: A law firm analyzing 100 contracts per day with Claude Opus 4.6 would spend $157.50/month. Switching to Gemini 2.5 Flash drops that to $4.50/month — a 35x reduction. The quality trade-off depends on how nuanced the analysis needs to be.
For document analysis, the input price is king. Models with low input rates (Gemini Flash-Lite at $0.075/M, GPT-5 nano at $0.05/M) dominate because the document is 16x larger than the response.
Task 6: Translation (500 Words)
Typical tokens: ~700 input (text + instructions), ~750 output
Translating a medium-length passage between languages. Token counts vary by language — Chinese and Japanese use more tokens per word than European languages.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.000042 | $0.000135 | $0.000177 |
| GPT-5 nano | $0.000035 | $0.0003 | $0.000335 |
| DeepSeek V3.2 | $0.000196 | $0.000315 | $0.000511 |
| Gemini 2.5 Flash | $0.000105 | $0.00045 | $0.000555 |
| GPT-5 mini | $0.000175 | $0.0015 | $0.001675 |
| GPT-5.2 | $0.001225 | $0.0105 | $0.011725 |
| Claude Sonnet 4.6 | $0.0021 | $0.01125 | $0.013350 |
Translation is balanced between input and output. For high-quality translation of marketing materials or legal documents, premium models justify the cost. For bulk translation of user-generated content, budget models work fine.
Task 7: Data Extraction (Structured Output from Text)
Typical tokens: ~1,000 input (raw text), ~200 output (JSON)
Extracting structured data from unstructured text — names, dates, amounts, addresses from invoices, receipts, or forms. The output is compact JSON, so output costs are minimal.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.00006 | $0.000036 | $0.000096 |
| GPT-5 nano | $0.00005 | $0.00008 | $0.00013 |
| Gemini 2.0 Flash | $0.0001 | $0.00008 | $0.00018 |
| GPT-5 mini | $0.00025 | $0.0004 | $0.00065 |
| DeepSeek V3.2 | $0.00028 | $0.000084 | $0.000364 |
| GPT-5.2 | $0.00175 | $0.0028 | $0.00455 |
| Claude Sonnet 4.6 | $0.003 | $0.003 | $0.006 |
✅ TL;DR: Data extraction is extremely cheap across all models. Even at 100,000 extractions per month, you'd spend $9.60 with Mistral Small or $600 with Claude Sonnet. Budget models handle structured extraction nearly as well as premium ones — this is where you should save money.
Task 8: Chatbot Conversation Turn (With History)
Typical tokens: ~2,000 input (system prompt + conversation history), ~400 output
A single turn in an ongoing conversation. The input grows with each turn as conversation history accumulates. This estimate assumes 5-6 previous exchanges in context.
| Model | Input Cost | Output Cost | Total Cost per Turn |
|---|---|---|---|
| GPT-5 nano | $0.0001 | $0.00016 | $0.00026 |
| Mistral Small 3.2 | $0.00012 | $0.000072 | $0.000192 |
| Gemini 2.5 Flash | $0.0003 | $0.00024 | $0.00054 |
| DeepSeek V3.2 | $0.00056 | $0.000168 | $0.000728 |
| GPT-5 mini | $0.0005 | $0.0008 | $0.0013 |
| GPT-5.2 | $0.0035 | $0.0056 | $0.0091 |
| Claude Sonnet 4.6 | $0.006 | $0.006 | $0.012 |
| Claude Opus 4.6 | $0.01 | $0.01 | $0.02 |
[stat] 10-50x How much chatbot costs multiply over a 20-turn conversation as context grows
The critical insight: chatbot costs accelerate with conversation length. Turn 1 might cost $0.001, but turn 20 costs $0.01+ because you're sending the entire conversation history each time. Implement conversation summarization or sliding window context to keep costs linear.
Task 9: RAG Query (Retrieval-Augmented Generation)
Typical tokens: ~4,000 input (query + retrieved chunks), ~500 output
A typical RAG pipeline retrieves 3-5 text chunks from a vector database and sends them alongside the user's question. This is the standard enterprise AI pattern.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.00024 | $0.00009 | $0.00033 |
| GPT-5 nano | $0.0002 | $0.0002 | $0.0004 |
| Gemini 2.5 Flash | $0.0006 | $0.0003 | $0.0009 |
| DeepSeek V3.2 | $0.00112 | $0.00021 | $0.00133 |
| GPT-5 mini | $0.001 | $0.001 | $0.002 |
| GPT-5.2 | $0.007 | $0.007 | $0.014 |
| Claude Sonnet 4.6 | $0.012 | $0.0075 | $0.0195 |
A knowledge base serving 50,000 queries per month:
- Mistral Small 3.2: $16.50/month
- GPT-5 mini: $100/month
- Claude Sonnet 4.6: $975/month
💡 Key Takeaway: For RAG, the retrieval quality matters more than the model. A cheap model with great retrieval beats an expensive model with poor retrieval every time. Invest in your embedding pipeline and chunking strategy before upgrading your generation model.
Task 10: Image Description / Vision Analysis
Typical tokens: ~1,000 input (image tokens + prompt), ~300 output
Describing an image or extracting information from a screenshot. Image tokens vary by resolution — a standard 1024x1024 image typically costs 1,000-1,500 tokens.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Gemini 2.0 Flash | $0.0001 | $0.00009 | $0.00019 |
| Gemini 2.5 Flash | $0.00015 | $0.00009 | $0.00024 |
| GPT-5 mini | $0.00025 | $0.0006 | $0.00085 |
| GPT-5.2 | $0.00175 | $0.0042 | $0.00595 |
| Claude Sonnet 4.6 | $0.003 | $0.0045 | $0.0075 |
| Claude Opus 4.6 | $0.005 | $0.0075 | $0.0125 |
| Gemini 3 Pro | $0.002 | $0.0036 | $0.0056 |
Vision tasks are moderately priced. For bulk image processing (product catalogs, document scanning), Gemini Flash models offer the best value by a wide margin.
Task 11: Code Review (Full File)
Typical tokens: ~3,000 input (code + review instructions), ~1,000 output
Reviewing a complete source file for bugs, style issues, and improvements. Output is longer than most tasks because the model provides detailed feedback.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| Mistral Small 3.2 | $0.00018 | $0.00018 | $0.00036 |
| DeepSeek V3.2 | $0.00084 | $0.00042 | $0.00126 |
| Codestral | $0.0009 | $0.0009 | $0.0018 |
| GPT-5 mini | $0.00075 | $0.002 | $0.00275 |
| GPT-5.2 | $0.00525 | $0.014 | $0.01925 |
| Claude Sonnet 4.6 | $0.009 | $0.015 | $0.024 |
| Claude Opus 4.6 | $0.015 | $0.025 | $0.04 |
⚠️ Warning: Code review quality varies dramatically between models. Budget models catch syntax errors and obvious bugs but miss architectural problems and subtle race conditions. For production code reviews, Claude Sonnet or GPT-5.2 are worth the premium — a missed bug in production costs far more than $0.02.
Task 12: Reasoning / Complex Analysis
Typical tokens: ~1,500 input (problem statement), ~2,000 output (step-by-step reasoning)
Complex tasks requiring multi-step reasoning — financial analysis, strategic planning, research synthesis. These benefit most from reasoning-class models.
| Model | Input Cost | Output Cost | Total Cost |
|---|---|---|---|
| o4-mini | $0.00165 | $0.0088 | $0.01045 |
| o3 | $0.003 | $0.016 | $0.019 |
| Gemini 2.5 Pro | $0.001875 | $0.02 | $0.021875 |
| GPT-5.2 | $0.002625 | $0.028 | $0.030625 |
| Claude Opus 4.6 | $0.0075 | $0.05 | $0.0575 |
| o3-pro | $0.03 | $0.16 | $0.19 |
| GPT-5.2 pro | $0.0315 | $0.336 | $0.3675 |
Reasoning tasks show the biggest price spread — 35x between o4-mini and GPT-5.2 pro. The premium models produce measurably better reasoning on hard problems, but for standard analytical tasks, o4-mini and Gemini 2.5 Pro deliver excellent results.
The Complete Cost-Per-Task Cheat Sheet
Here's every task with the cheapest viable option and the premium recommendation:
| Task | Cheapest Option | Cost | Premium Pick | Cost |
|---|---|---|---|---|
| Email summary | Mistral Small 3.2 | $0.00003 | GPT-5 mini | $0.00025 |
| Support response | Mistral Small 3.2 | $0.000084 | Claude Sonnet 4.6 | $0.006 |
| Blog post | Mistral Small 3.2 | $0.0004 | Claude Sonnet 4.6 | $0.031 |
| Code generation | Mistral Small 3.2 | $0.00017 | Claude Sonnet 4.6 | $0.013 |
| Document analysis | Mistral Small 3.2 | $0.00057 | Gemini 3 Pro | $0.022 |
| Translation | Mistral Small 3.2 | $0.00018 | GPT-5.2 | $0.012 |
| Data extraction | Mistral Small 3.2 | $0.00010 | GPT-5 mini | $0.00065 |
| Chatbot turn | Mistral Small 3.2 | $0.00019 | Claude Sonnet 4.6 | $0.012 |
| RAG query | Mistral Small 3.2 | $0.00033 | GPT-5.2 | $0.014 |
| Vision analysis | Gemini 2.0 Flash | $0.00019 | Claude Sonnet 4.6 | $0.0075 |
| Code review | Mistral Small 3.2 | $0.00036 | Claude Sonnet 4.6 | $0.024 |
| Complex reasoning | o4-mini | $0.01 | GPT-5.2 pro | $0.37 |
✅ TL;DR: Mistral Small 3.2 is the cost king for almost every task at $0.06/$0.18 per million tokens. For quality-sensitive tasks (code review, complex reasoning, customer-facing content), Claude Sonnet 4.6 and GPT-5.2 justify their 50-100x premium. Use our calculator to model your exact usage mix.
Frequently asked questions
How much does a single AI API call cost?
Most AI API calls cost between $0.00003 and $0.05 depending on the task and model. Simple tasks like email summarization cost fractions of a cent even with premium models. Complex reasoning tasks with large outputs can reach a few cents per call. Use our calculator to estimate costs for your specific workload.
Which AI model gives the best value for money?
Mistral Small 3.2 offers the best raw cost efficiency at $0.06/$0.18 per million tokens — it's the cheapest option for 10 out of 12 common tasks. For balanced quality and cost, GPT-5 mini ($0.25/$2.00) and Gemini 2.5 Flash ($0.15/$0.60) are strong mid-tier choices. See our cheapest AI APIs guide for the full breakdown.
How do I reduce my AI API costs?
The most impactful strategy is model routing — use cheap models for simple tasks and expensive models only when quality matters. Beyond that: implement prompt caching, use batch APIs for non-urgent tasks, compress prompts to reduce input tokens, and set max token limits on outputs. Companies typically cut costs 40-70% with these techniques. Check our 10 strategies guide for implementation details.
Does using a more expensive AI model always give better results?
No. For structured tasks like data extraction, classification, and simple Q&A, budget models perform within 5% of premium models. The quality gap widens for creative writing, nuanced reasoning, and complex code generation. Test your specific use case with a cheap model first — you might be surprised. Our token guide explains how to evaluate whether you're overpaying.
How much does it cost to run an AI chatbot for my business?
A chatbot handling 1,000 conversations per day (average 10 turns each) costs between $2/day with GPT-5 nano and $200/day with Claude Opus 4.6. Most businesses land around $10-30/day using mid-tier models like GPT-5 mini or Gemini 2.5 Flash. The AI chatbot cost breakdown post has detailed projections for different scales.