AI Rate CardHow much does a coding agent cost a month? A calculator with live prices for 13 models
Updated Nov 9, 2026
A coding agent is expensive in an unexpected way: not because of the code it writes, but because it resends its whole growing context - files, tool results, the conversation - with every call. Thirty calls of 30,000 tokens is almost a million input tokens for one task. The coding agent calculator prices that month on 13 usual coding models at once, at the cheapest real-time seller of each, with each model's reasoning class beside the price.
How to use it
- Tasks a day: how many jobs you give the agent (a bug fix, a feature, a review).
- Model calls per task: an agent step is one call - read a file, run tests, edit. 20-50 is common.
- Context per call: tokens sent each time; it grows during a task, so take the average (20,000-60,000).
- Answer and reasoning per call: both are billed as output; reasoning models think before they answer.
- Share from the cache: agents resend mostly the same context, so 60-90% is typical where the seller supports caching.
- Models: leave empty for the usual 13, or list your own ids.
Example from today's prices
Default workload - 10 tasks a day, 30 calls, 30,000 tokens of context, 2,300 output tokens per call, 70% cached - is 270M input and 21M output tokens a month:
- Claude Haiku 5.5 about $10, DeepSeek V4.1 Flash about $17, GLM-5.3 about $33 a month;
- Gemini 3.8 Flash about $67, Kimi K3 about $82;
- Claude Sonnet 5.5 about $194, Gemini 3.1 Pro about $224, Claude Opus 5.5 about $388, GPT-5.5 about $653. Cheap is not the same as good: check the class column - an S or A model that finishes the task in fewer calls can cost less overall.
What it leaves out
- Cache write fees, tool and search fees, and retries after errors.
- A subscription (Claude Pro/Max, Cursor plans) may be cheaper for heavy personal use - see subscription or API.