Cost by model
Figures are USD per 1M tokens, as of 2026-09. The rows are GPT, Claude, GLM, DeepSeek, Qwen, and a local run. This sheet has no Gemini price. Anthropic's Claude rows match the vendor's published list prices; the other API rows follow the sheet this calculator uses. Re-check the vendor before you budget. Where a vendor publishes a cache-read rate, the calculator uses it. Otherwise it falls back to a flat 50% discount on input, which is a placeholder rather than a measured figure.
| Model | Input / 1M | Output / 1M | Cache |
|---|---|---|---|
| GPT-6.1 Sol | $2 | $10 | Yes |
| GPT-6 Astra | $10 | $50 | Yes |
| Claude Opus 5.5 | $4 | $20 | Yes |
| Claude Fable 5.1 | $10 | $50 | Yes |
| Claude Sonnet 4.8 | $3 | $15 | Yes |
| GLM-5.3 | $0.6 | $2.2 | Yes |
| DeepSeek-V4-Pro | $0.27 | $1.1 | Yes |
| Qwen3.8-27B (API) | $0.4 | $1.6 | No |
| Local deploy (Qwen/GLM) | $0 | $0 (runs on a 24GB GPU or 32GB Mac) | No |
Last updated: 2026-09.
What drives the cost
Five things move the month. Token price is the rate in the table above. Task complexity is how many input and output tokens one loop carries: the support preset is 2,000 in and 400 out, the research preset is 30,000 in and 2,000 out, and the coding preset is 60,000 in and 6,000 out. Tool calls show up as extra rounds, because the same context is resent each time the agent looks something up. Hosting the weights and any vector store are not in the token math; local weights are $0 in tokens and still need a machine. Human maintenance — reviewing bad runs, tightening tools, watching the bill — is time, and this page does not convert it into dollars.
Cost per task
Cost per task is one completed loop at 3 rounds with the cache discount, not a single chatbot reply. A support task is about $0.002 on DeepSeek-V4-Pro and about $0.07 on GPT-6 Astra. A coding task is about $0.04 to $1.08 on those same two models. If the agent retries, raise the round count. Do not pretend the retry was free.
Try the calculator
Pick a model, a daily task count, and how many rounds a task usually takes. The result is a month of inference at your volume. It runs in the browser. Nothing you type is stored.
Estimate monthly agent spend
Results update as you type. Nothing is submitted or stored. Last updated: 2026-09.
Estimated monthly cost
$18.9 / month
Cost per task
$0.0126
Daily cost
$0.63
Same workload, all 7 models
- GPT-6.1 Sol$18.9
- GPT-6 Astra$99
- Claude Opus 5.5$37.8
- Claude Fable 5.1$92.25
- Claude Sonnet 4.8$40.5
- GLM-5.3$6.66
- DeepSeek-V4-Pro$3.2
- Qwen3.8-27B (API)$6.48
- Local deploy (Qwen/GLM)$0 · runs on a 24GB GPU or 32GB Mac
Switching to DeepSeek-V4-Pro would cut your bill by 83%.
Want to run agents on your own GPU? See Run Qwen3.8-27B locally · RunPod · Vast.ai
Monthly cost is 30 × daily spend. Cost per task is one loop at the tokens and rounds you set. The local row is $0 in API tokens.
Cost by use case
Same volume for each: 50 tasks a day, 3 rounds, cache discount, 2026-09 prices. The range is DeepSeek-V4-Pro through GPT-6 Astra.
- Support agent. Short context. About $3 to $99 per month, or about $0.002 to $0.07 per task.
- Research agent. A long page plus a summary. About $28 to $585 per month.
- Coding agent. A large codebase in the prompt. About $66 to $1,620 per month, or about $0.04 to $1.08 per task.
How to cut the bill
- Cache the repeated prompt. Turn the switch on if the model supports it. This page uses the vendor's cache-read rate when it has one, and a 50% input haircut when it does not.
- Route the easy step to a smaller model. OrcaRouter is one way to send the cheap turns off the frontier model.
- Batch the work that can wait. Many providers cut the rate for async jobs.
- Run the weights locally when the token line is the whole problem. Qwen3.8-27B locally is $0 in token fees.
- Cap tokens and rounds. A retry limit is a budget. The token budget calculator is the place to set that ceiling before the month runs away.
FAQ
How much AI agent cost should you plan for?
On this page's sheet, at 50 tasks a day and 3 rounds with a cache discount, a short support loop is about $3 to $99 per month. A coding loop on the same volume is about $66 to $1,620 per month. The low end is DeepSeek-V4-Pro. The high end is GPT-6 Astra. Local weights are $0 in tokens.
What is the cost of an AI agent?
The cost of an AI agent on this sheet is inference: input tokens, output tokens, and how many times the loop resends them. A hosted product's seat fee is a separate bill this table does not price. Add engineering time on top if you are building the agent rather than calling an API.
What is the cost of AI agents when several run at once?
The cost of AI agents scales with how many loops you run, not with a shared discount. Two agents at the same volume cost about twice one agent, unless they share a cached prompt. Use the calculator twice and add the months.
What is the cost per task of an AI agent?
At 3 rounds with the cache discount, one support task is about $0.002 on DeepSeek-V4-Pro and about $0.07 on GPT-6 Astra. One coding task is about $0.04 to $1.08 on the same models. Cost per task is one completed loop, not one chat reply.
Is an AI agent expensive?
A short support agent on a cheap API is often a few dollars a month at 50 tasks a day. The same volume as a coding agent on GPT-6 Astra is about $1,620 a month. Expensive is the long context resent on every round, not the fact that the program is an agent.
The loop those prices sit on is how do AI agents work. The map of types and tools is the AI agent guide. The build is how to create an AI agent.