AI agent workflows: what it costs and which AI to use
An agent that plans, calls tools and loops until a task is done: research, triage, data clean-up, back-office jobs. Teams that automate multi-step work and are surprised when the invoice is ten times the demo.
The short answer
At 1,500 runs a month, we estimate $7.44 on DeepSeek V4 Flash 0423 (the cheapest model that cleared our quality bar), $71.67 on Kimi K2.5, and $735 on GPT-5.5. The flagship costs 10x the middle pick. Judged on tool-calling, reasoning and code-fix tasks: picking the right tool with valid arguments, every step.
Simulate your AI agent workflows
Pick a job, set how much of it you do, and see what every model would cost. Defaults are typical, not yours. Change them.
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.
1,500 runs a month, three ways to pay
Every model here cleared a score of 9.4 out of 10 on our tool call, reasoning, code fix benchmark tasks, ran without errors and fits a 128K-token window. 30 of 386 priced models qualify.
Cheapest model that clears the bar.
DeepSeek V4 Flash 0423
- Per run
- $0.0050
- Benchmark
- 10.00 / 10 · 9 prompts
- Context
- 1.05M tokens
Priced near the middle of the models that clear the bar; best score there.
Kimi K2.5
- Per run
- $0.048
- Benchmark
- 10.00 / 10 · 9 prompts
- Context
- 262K tokens
What many people reach for first.
GPT-5.5
- Per run
- $0.490
- Benchmark
- 10.00 / 10 · 9 prompts
- Context
- 1.05M tokens
Choosing GPT-5.5 over Kimi K2.5 costs $663 more a month, $7,958 a year (10x the price).
Both scored 10.00 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 · 1,500 | 3x · 4,500 | 10x · 15K | 100x · 150K |
|---|---|---|---|---|
| DeepSeek V4 Flash 0423 | $7.44 | $22.32 | $74.41 | $744 |
| Kimi K2.5 | $71.67 | $215 | $717 | $7,167 |
| GPT-5.5 | $735 | $2,205 | $7,349 | $73,485 |
What a budget buys
- DeepSeek V4 Flash 042320Kruns a month, about 671 a day
- Kimi K2.52,092runs a month, about 69 a day
- GPT-5.5204runs a month, about 6 a day
Every model, priced for your workload
41 models, 45 hidden because their window is smaller than 128K tokens.
| Model | A month | Per run | Benchmark | Context |
|---|---|---|---|---|
| DeepSeek V4 Flash 0423 DeepSeekBudget pick | $7.44 | $0.0050 | 10.00 (9) | 1.05M |
| Ministral 3 8B 2512 Mistral AIbelow the bar | $8.46 | $0.0056 | 7.00 (9) | 262K |
| DeepSeek V4.1 Flash DeepSeek | $11.76 | $0.0078 | 10.00 (9) | 1.05M |
| Llama 4 Scout Metabelow the bar | $12.52 | $0.0083 | 8.44 (9) | 1.31M |
| Qwen3.8 Flash Alibaba | $14.33 | $0.0096 | 10.00 (9) | 1M |
| GLM 5.3 Flash Z Aibelow the bar | $15.89 | $0.011 | 9.00 (9) | 1.05M |
| Mistral Small 4 Mistral AI | $16.61 | $0.011 | 10.00 (9) | 262K |
| MiniMax M2.5 Minimax | $27.76 | $0.019 | 10.00 (9) | 205K |
| GPT-5.4 Nano OpenAI | $30.30 | $0.020 | 10.00 (9) | 400K |
| GLM 5.3 Z Ai | $32.52 | $0.022 | 10.00 (9) | 1.05M |
| Muse Glimmer 30B Meta | $33.95 | $0.023 | 10.00 (9) | 131K |
| MiniMax M2.7 Minimax | $35.40 | $0.024 | 10.00 (9) | 205K |
Benchmark = mean judge score out of 10 on the tool call, reasoning, code fix 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.
What we assumed
- Fifteen hundred runs a month.
- A run is about 15 model calls. Across them, 60,000 input tokens are billed (context re-sent each step) and 7,000 tokens are written.
- 70% of input repeats from step to step and can be cached where supported.
- Reasoning adds 50% to billed output. Retries and tool errors add 15% more calls.
These are typical numbers, not measurements of your business. Edit them in the simulator above.
What drives the cost
Context snowballs. Each step re-sends everything before it, so a 20-step run can send the same instructions twenty times.
Watch out for
- Agent cost is the least predictable of all the use cases. Cap steps per run and set a hard monthly spend limit.
- Without caching, the same run can cost three to five times as much on models that do not discount repeated input.
- The benchmark scores single steps, not whole runs. Longer chains fail more often than any single step suggests.
Questions people ask
How much will AI agent workflows cost per month?
At 1,500 runs a month, we estimate $7.44 a month on DeepSeek V4 Flash 0423, the cheapest model that clears our quality bar for this job. Kimi K2.5 costs $71.67 and GPT-5.5 costs $735. These are estimates from the assumptions listed on this page, not a quote. Change the numbers to match your own volume.
What is the cheapest AI model for AI agent workflows?
On current OpenRouter prices, DeepSeek V4 Flash 0423 from DeepSeek is the cheapest model that cleared our bar of 9.4 out of 10, scoring 10.0/10 on 9 benchmark prompts for this kind of work. The benchmark is small, so treat it as a shortlist and test on your own data before switching.
Is it worth paying for a flagship model for AI agent workflows?
On our benchmark, GPT-5.5 scored 10.0/10 on this kind of work, and a model priced near the middle of the models that cleared the bar, Kimi K2.5, scored 10.0/10, while the flagship costs 10x as much. A flagship buys safety margin the benchmark cannot measure. If a wrong answer is expensive for you, that can be worth paying for. If not, you are probably overpaying.
What happens to the cost if AI agent workflows grows ten times?
Pay-per-use pricing scales in a straight line, so ten times the runs is ten times the bill. Use the growth table on this page to see 3x, 10x and 100x for the same three models.
What makes AI agent workflows expensive?
Context snowballs. Each step re-sends everything before it, so a 20-step run can send the same instructions twenty times.
Price a different job
- Customer support chatbotFrom $1.74 a month at 3,000 conversations
- Coding assistant for a teamFrom $4.12 a month at 4,400 requests
- Document summarisationFrom $1.86 a month at 2,000 documents
- Blog and content writingFrom $0.083 a month at 120 articles
- Knowledge base Q&A (RAG)From $5.62 a month at 15K questions
- Invoice and form data extractionFrom $4.65 a month at 20K documents
- Ticket and content classificationFrom $11.30 a month at 500K items
- Sales email personalisationFrom $0.595 a month at 5,000 emails
- Meeting notes and summariesFrom $0.575 a month at 400 meetings
- Code review for pull requestsFrom $0.840 a month at 600 pull requests
- Personal AI assistant for a founderFrom $0.478 a month at 1,500 messages
- Text-to-SQL data analystFrom $2.54 a month at 6,000 queries