LocalOrCloud

What will your AI agent really cost each month?

Agents reread your whole project again and again, so they burn far more tokens than a chat. Pick your agent, your model and how hard you work it. You get a monthly cost range and the cheapest plan for that workload.

Preview · every price checked · sources at the bottom

Your agent

Type of agent

Claude Code, Copilot, Cursor or your own agent on an API.

Model

Hours of agent work per day

Days per month

How hard do you work it?

Cache hit rate: Share of input the provider has already seen and bills at the cheap cached rate. Well-tuned agents reach 80% or more.

Prompts often go over 100K tokens Big repos and long sessions do. Only Claude Haiku 5.5 charges more for this.

Edit the token assumptions

Input tokens per hour

Output tokens per hour

One study measured about 4 million tokens for an agent to fix a single bug in a real repository. Most of that is the agent rereading files.

Your estimate on pay-as-you-go API pricing

Your options for this workload

Option	Monthly price	What you'd pay

Copilot bills agent work at each model's API rate once its included usage runs out, so we can calculate it. Claude Code and Cursor cap usage in time windows instead of dollars. If your API cost is far above their price, a plan usually saves money, as long as its limits don't stop your work.

Could your own GPU do this for $0 in tokens?

Our next tool checks which local models your graphics card or Mac can run, then compares that with the bill above. It's coming soon.

Prices used and where they come from

Model	Input / 1M	Cached / 1M	Output / 1M	Source

Plan	Price / month	Included usage	Source

Cached prices marked "est." assume 10% of the input price. The range runs from 0.6× to 1.8× the estimate, because the same task can cost very different amounts from one run to the next. This is a planning estimate, not a bill. Always confirm on the provider's own pricing page. Cursor moved to new plans in October 2026; legacy Pro plans still work for existing customers.