A seat count times a sticker price is not a budget. This builds the forecast finance will actually review: team by team, month by month, prepaid capacity separated from the metered bill, every assumption named, and the whole thing exportable as a CSV or a memo.
Hiring is your approved plan and applies to every archetype evenly; 3 percent a month is a placeholder, so replace it with the number your headcount plan actually carries. Usage drift is an assumption, applied to the metered half only, because seat spend is flat until somebody changes tier. Agent usage compounds as teams run more agents per person and longer runs per agent, but no vendor publishes a rate, so 5 percent a month is a planning figure rather than a measurement. Set it to zero to forecast a floor.
Headcount is the sum of the three archetype columns. Light wants completions and the odd question, regular runs an agent for part of most days, heavy runs agents most of the day. API % is metered overflow expressed as a share of that team's seat spend; the assumptions panel prints what it comes to per developer so you can check it against a real invoice. Leave a budget at zero and the variance column stays blank.
Add a team to see the forecast.
The forecast above is only defensible because it keeps prepaid capacity, metered usage and growth apart. Here is the method behind it, in the order a finance review will ask.
The most useful outside figure available is Anthropic's, because it was measured rather than modelled: across enterprise deployments, Claude Code averages about $13 per developer per active day and $150 to $250 per developer per month on metered billing, with 90 percent of users staying under $30 per active day. Read it as a ceiling on flat plans and a centre on metered ones. It beats any vendor's headline seat price as an anchor, because it is what the work costs rather than what a licence costs.
Use it as a sanity check on the number this page produces, not as an input. If your blended per-developer figure lands at $400, either your archetype mix is wrong or somebody is on a tier they do not need.
Three archetypes cover almost every engineering organisation, and the split is usually closer to 40 / 40 / 20 than people expect. The reason to split this way rather than by team or seniority is that the archetype predicts consumption and nothing else does. A staff engineer who reviews more than they write costs less than a mid-level engineer running three agents in parallel. Pricing the archetype instead of the headcount is what stops a light user sitting on a $200 plan, which is the single most expensive mistake at scale.
The presets in the planner are real stacks at real list prices, and section 03 prints exactly which seats each one buys. Pick the preset closest to what a team is on rather than the one you wish they were on. The forecast is only useful if it describes today.
On a flat plan tokens are prepaid: one more turn costs nothing, and unused capacity is not a refund, it is waste. On an API key or a credit pool every token is charged, and one unattended run can move the monthly number materially. That is why this planner applies hiring growth to both lines but usage drift to the metered line only. Seat spend does not compound. It steps, and only when somebody changes tier.
Keeping them apart also assigns the variance. A seat overage is a headcount conversation with a hiring manager. A metered overage is a usage conversation with an engineering lead, and it is answerable with spend caps and model policy rather than with a renegotiated contract.
Two numbers in any AI budget are guesses: how fast the team grows and how fast usage per person grows. The first is knowable, because hiring is planned and approved. The second is not, because no vendor publishes a growth rate and your own history is short. Writing 5 percent a month and labelling it an assumption is honest. Writing 5 percent a month and calling it research is how a budget loses its credibility in the first review.
Then put an alert threshold on it. A budget with no threshold is a number somebody discovers was wrong at the end of the quarter. A budget with a threshold at 90 percent tells you which team crosses first and in which month, which is enough lead time to move a tier or cap a key before the invoice arrives. The verdict line above names that month for you.
Every figure this page produces came out of a form, including the two assumptions. Before the next renewal, replace them: per-seat activity data settles how many licences are idle, and a real per-developer spend reading settles the drift rate. Until then, forecast honestly and label the parts you are guessing at.
Related: AI spend management covers the controls that keep a forecast true, and the AI team cost calculator prices a single month in more detail if you want to interrogate one team's stack before it lands in this forecast. Browse all free AI cost and governance tools for the adjacent cap, policy, and review workflows.
Every seat price below was read from the vendor's own pricing page and checked August 2026. These are the exact figures the forecast multiplies, so the arithmetic is auditable without opening the console.
| Preset | Light seat | Regular seat | Heavy seat |
|---|---|---|---|
| Copilot Business plus Claude | Copilot Business $19 | Copilot Business $19 plus Claude Team standard $25 | Copilot Business $19 plus Claude Max 20x $200 |
| Claude first | Claude Pro $20 | Claude Team standard $25 | Claude Max 20x $200 |
| Copilot only | Copilot Business $19 | Copilot Business $19 | Copilot Max $100 |
| Cursor plus ChatGPT | Cursor Pro $20 | Cursor Teams $40 | Cursor Teams $40 plus ChatGPT Pro $200 |
| ChatGPT and Codex | ChatGPT Business $25 | ChatGPT Business $25 | ChatGPT Pro $200 |
| Gemini Code Assist | Standard $22.80 | Standard $22.80 | Enterprise $54 |
| Continuum | Free with your own keys | Continuum Plus $25 | Continuum Max 200 $200 |
| Custom | Three prices you type per team. Nothing is assumed. | ||
Prices are per seat per month on monthly billing unless stated. Claude Team standard is $25 monthly and $20 annual; ChatGPT Business is the same pair, which is the closest thing to a like-for-like seat comparison on the market. Gemini Code Assist has had no free individual tier since June 2026. Continuum's app is free with your own keys; the paid tiers add hosted inference with a weekly budget.
Deeper on controls in AI spend management.
Build it from three separate lines rather than one blended number. First, prepaid seat capacity: headcount by usage archetype multiplied by the list price of the tier each archetype actually needs. Second, metered API overflow: the part of the bill that moves with usage and is the only line that can surprise you. Third, growth: approved hiring plus a stated usage-drift assumption for the metered half. Keep the three visible in the forecast, because a variance in seat spend is a headcount question and a variance in metered spend is a usage question, and they have different owners.
Anthropic's own enterprise documentation puts metered Claude Code usage at about $13 per developer per active day and $150 to $250 per developer per month, with 90 percent of users under $30 per active day. On flat subscriptions the same work usually lands between $20 and $200 per developer per month depending on tier. A defensible planning number for a team that codes with agents daily is $100 to $200 per developer per month, and materially less for people who only want completions.
Nobody publishes a credible figure, so treat any number here as an assumption you are choosing rather than a measurement you are citing. Agent usage compounds for a structural reason: teams add more agents per person, runs get longer as trust grows, and the work that was manual last quarter becomes an agent task this quarter. This planner defaults to 5 percent per month on the metered half only, because seat spend is flat until you change tiers. Set it to zero if you would rather forecast a floor and treat every overage as news.
Because only one of them can run away. Seat spend is prepaid capacity: one more agent turn costs nothing, and unused capacity is waste rather than a refund. Metered API spend is charged per token, so a single unattended job can move the month. A budget that shows one combined figure cannot answer the question finance asks in month three, which is which half of the variance is at risk and who owns it. Continuum for organizations puts a weekly spend cap and a model policy behind the metered half, per team and per person.
Two numbers in the memo you just exported are assumptions, and both are answerable with data you already have. The scan reads the AI history on a developer's machine and prices it in about sixty seconds, which replaces the drift guess with a reading. Continuum for organizations puts weekly spend caps and model policy behind the metered half, so the forecast stops being a hope.
one command · runs locally · free and standalone