How to cost and cap AI coding sessions before your team starts using them

Price the session before the team touches it

Before your team starts using AI coding agents, decide what one session will cost and how much it can spend. Linear moved coding sessions to a lower-cost, two-part pricing model on August 20, 2026, where model usage and runtime determine AI credit cost. Set the caps before rollout, not after the surprise invoice.

Linear Agent prepares, executes, and checks code before handing the result back to users. During a session, it can find and add the toolchains and dependencies a task needs. The sessions ship with Python, Ruby, Go, and other languages ready to use. In configurable environments, the agent can also pin runtime versions and add setup scripts.

For a small team, the practical formula is simple: session cost = (input tokens × input rate + output tokens × output rate) + (sandbox blocks × runtime rate). If the provider publishes cache-write and cache-read rates, add those lines. If your provider bills sandbox time in blocks, use the published block rate in the formula. The token side is the easy part. For Claude Opus 5, a third-party model, Cursor publishes rates at $5 per million input tokens, $25 per million output tokens, $6.25 per million cache-write tokens, and $0.50 per million cache-read tokens. On Teams and Enterprise plans, Cursor charges an extra $0.25 per million tokens for third-party model requests, above the model API price.

Do not treat the model rate as the whole AI coding cost. Sandbox runtime is the part that grows when an agent keeps retrying, installing dependencies, or running tests. The formula should make that growth visible before it shows up on the invoice.

Set the caps before enabling sessions

Linear gives admins workspace-wide and per-user caps, with resets that run daily, weekly, or monthly, and AI credit use pauses when a cap is hit. That pause is the control you want in a small company: no one needs to be on call for an uncontrolled session.

Cursor ships usage meters and spend alerts alongside its pricing changes. Cursor's Teams plans add billing and administration controls, plus usage analytics. Cursor's Pro plan comes with $20 of third-party model usage each month; Pro Plus gets $70, and Ultra gets $400. Cursor splits included usage into two pools, each resetting on the monthly billing cycle. The included pools matter because a cap that ignores them misses the spend the team already paid for. If your team is on a plan with included usage, the cap should cover the included pool and any overage.

Letting a few power users spend the pool before the rest of the team has a baseline is the mistake. Put the caps in place before enabling sessions.

Run early usage like a controlled test

The checklist below turns the formula and caps into a repeatable ritual.

• Collect the provider rates. The sheet lists input, output, cache-write, cache-read, and any platform surcharge.
• Add the sandbox runtime line. The estimate is incomplete until it includes the runtime portion.
• Set the caps. The written limit covers the workspace and each user, with a reset cadence and a named owner.
• Enable meters and alerts. The dashboard shows spend by user and model before anyone hits a hard stop.
• Review early usage. The note lists the top users, models, and charge types.

Finout can pull Cursor spend through the Admin API and split it by user, model, charge type, and day. You do not need a full finance stack for early usage. You need enough detail to answer the key questions: who used the most, which model drove the cost, and whether the caps were too tight or too loose.

Store the session formula in the same document as the caps. When a new model or a new runtime option appears, update the rate, rerun the estimate, and adjust the limit before the team sees the new option.