How much should we budget for AI coding agents per developer in 2027?
Budget by usage tier, not one number per head. At Claude Sonnet 5.5 list prices, a light agent user costs about $37 a month, a typical one $73 and a heavy one $591. With 3 light, 5 typical and 2 heavy developers in every 10, that averages $166 per developer a month, $1,991 a year, before a reserve for growth and price changes. Opus 5.5 doubles it. Seats at $19 to $125 a developer look cheaper until heavy users hit limits and pay overage, the API bills exactly what you use so the total moves, and a fixed budget per developer caps each line.
Start from usage tiers
Agent spend is lopsided. On the Standard Compute gateway from September 14 to October 13, 2026, the median active coding-agent day was 95 requests and 8.6M tokens, and a busy day (the 90th percentile) 491 requests and 87.0M tokens. Priced at list rates, a busy day costs about 8 times a typical one, and 87% of input tokens were cache reads, which keeps both far below their raw token count.
So sort developers into three tiers before you multiply by headcount. Light users run an agent on about half their working days, typical users every day, and heavy users run long or parallel sessions every day.
| Tier | Assumed use | Claude Sonnet 5.5 | Claude Opus 5.5 | GPT-6 Luna |
|---|---|---|---|---|
| Light | Median day, 10 days a month | $37 | $73 | $2 |
| Typical | Median day, 20 days a month | $73 | $147 | $5 |
| Heavy | Busy day, 20 days a month | $591 | $1,182 | $49 |
Monthly cost per developer at list prices: median day $3.67 on Sonnet 5.5, $7.34 on Opus 5.5 and $0.23 on GPT-6 Luna; busy day $29.55, $59.10 and $2.43. The sample is agents that name themselves in requests, such as OpenCode, Kilo Code, Roo Code and GitHub Copilot Chat; Claude Code and Codex may differ. A busy day every working day is an upper bound.
The budget worksheet
Copy the lines into your own sheet. The multipliers are Claude Sonnet 5.5 list prices (use the tier table above for other models), and the example columns use Sonnet 5.5 for every request and the 3 / 5 / 2 mix per 10 developers, which is our planning assumption, not a measurement of your team.
| Line | How to fill it | 10 developers | 50 developers |
|---|---|---|---|
| 1. Light users | Headcount × $36.70 | 3 people: $110 | 15 people: $551 |
| 2. Typical users | Headcount × $73.40 | 5 people: $367 | 25 people: $1,835 |
| 3. Heavy users | Headcount × $591.00 | 2 people: $1,182 | 10 people: $5,910 |
| 4. Usage per month | Lines 1 to 3 | $1,659 | $8,296 |
| 5. Model mix | × 2 if Opus 5.5 is the default; far less if routine steps go to efficient models | × 1 (Sonnet 5.5) | × 1 (Sonnet 5.5) |
| 6. Seats or platform fees | From the approach you pick (below) | $0 on the API | $0 on the API |
| 7. Reserve | 20% of line 4 × line 5 for new hires, more autonomy and price changes | $332 | $1,659 |
| 8. 2027 budget | (Line 4 × line 5 + line 6 + line 7) × 12 | $23,891 | $119,455 |
The 20% reserve is our suggestion for usage-billed lines. Each developer who moves from typical to heavy adds $518 a month on Sonnet 5.5 and $1,035 on Opus 5.5, so the heavy headcount is the line to get right.
Approach 1: per-seat subscriptions
A seat plan is the lowest price per developer on paper, and it bundles the vendor's editor or chat features. What it does not fix is agent usage: every plan in the table below either limits it or bills past an included amount.
- For: a known seat invoice, the lowest entry price, and IDE features such as completions that agent APIs don't replace.
- Against: heavy agent users hit 5-hour, weekly or credit limits, and the overage is billed at API rates, so the real cost of a heavy developer is close to the API cost plus the seat. Copilot Business works out to roughly the larger of the seat and the usage, because the seat is the credit pool.
- Against: one vendor's models and tools. Vendors change what a seat includes mid-contract; Copilot moved to AI credits on June 1 and retires three models on November 16.
Seat prices to plan with
| Plan | Seat price | Agent usage included | Past the limit |
|---|---|---|---|
| Claude Team, Standard seat | $20/mo yearly, $25 monthly | Claude Code included; limits per 5-hour session, plus weekly or monthly caps | Wait, upgrade, or extra usage at standard API rates |
| Claude Team, Premium seat | $100/mo yearly, $125 monthly | 5× a Standard seat's usage | Extra usage at standard API rates |
| Claude Enterprise | $20/mo, billed yearly, 20-seat minimum | None: usage is billed at API rates on top | Admin spend limits per user and organization |
| ChatGPT Business | $20/mo yearly, $25 monthly | Codex, with limits that depend on model and task | Buy workspace credits |
| GitHub Copilot Business | $19/user | 1,900 AI credits ($19) per user, pooled; completions use none | $0.01 per credit, on by default |
| GitHub Copilot Enterprise | $39/user | 3,900 AI credits ($39) per user, pooled | $0.01 per credit, on by default |
| Cursor Teams | From $30/user | A pooled Cursor usage allowance | Claude, GPT and other third-party models billed at API list price |
| Kilo Teams | $15/user | None: team features only | Model usage at provider rates (Kilo Gateway, no markup) or your own keys |
| Windsurf, now Devin Desktop Teams | $80/mo per team + $40 per developer | A quota per seat that refreshes daily and weekly | Extra usage at API pricing |
Read on each vendor's pricing page or docs on October 16, 2026. Copilot's AI credits replaced premium requests on June 1, 2026, and Copilot prices Claude Sonnet 5.5 at $2 / $10 per million tokens, the same as Anthropic's list price. Windsurf's pricing page now redirects to Devin's.
Approach 2: pay per token through the API
Every agent works with an API key, any model is available, and you pay for exactly what is used: $1,659 a month for the example team of 10 on Sonnet 5.5, $8,296 for 50. That is the most accurate bill and the least predictable one.
- For: no seat fees, no session or weekly limits, and the freedom to move each task to a cheaper model.
- Against: the total is a forecast. One more heavy developer adds $518 a month on Sonnet 5.5, a switch to Opus 5.5 doubles the line, and a runaway autonomous session is billed in full.
- Against: in Anthropic's console, spend limits sit at the organization or workspace level, so a cap stops everyone in it at once; a limit per developer needs a workspace or key setup per person, or a proxy you run yourself.
Approach 3: a fixed budget per developer, with routing
Each developer gets their own API key with a monthly plan, and the plan's budget is the most that key can cost. This is how Standard Compute works.
- Cost: the router sends each step to the cheapest model that can do it, and providers are vetted for price, uptime and speed, with automatic fallback when one fails. Routine steps on efficient models are what stretch a budget past what the same work costs on one frontier model.
- Control: opt-in Compliance rules keep inference in Europe or the United States for the whole company, a team or one key; a team or key can narrow the company rule but never widen it. Restricted keys only reach provider endpoints with no training and zero data retention.
- Privacy: we do not use prompts or responses to train models, and request logs hold metadata, not prompt text.
- Billing: plans per key from $19 a month, Business plans from $499 (a $549 budget). A key at its budget stops with HTTP 402 and never bills overage. Developers you can't size yet use pay as you go, usage plus 15% from a prepaid company balance with a monthly limit per key. Each plan key is invoiced monthly to the company, with its address and VAT number.
- Against: a key that runs out stops until renewal or an upgrade, so an undersized plan blocks a sprint. A key pinned to Sonnet 5.5 all day still draws at its per-token rate: Pro Plus covers about 18 busy Sonnet 5.5 days on its own. Copilot's cloud agent and Cursor's own models can't use the key, and we have no SOC 2 report yet.
Worked example: 10 and 50 developers
| Approach | 10 devs, month | 10 devs, year | 50 devs, month | 50 devs, year | What moves it |
|---|---|---|---|---|---|
| Claude Team seats | from $360 | from $4,320 | from $1,800 | from $21,600 | Light and typical on Standard, heavy on Premium seats, billed yearly. Extra usage at API rates whenever a seat hits its limit |
| Copilot Business | $1,659 | $19,909 | $8,296 | $99,546 | Agent usage as in the worksheet. Seats ($190 and $950) become pooled credits; the rest bills at $0.01 a credit |
| Kilo Teams + usage | $1,809 | $21,709 | $9,046 | $108,546 | $15 seats plus model usage at provider rates, so it moves like the API |
| API, Sonnet 5.5 | $1,659 | $19,909 | $8,296 | $99,546 | Each typical-to-heavy switch +$518 a month; Opus 5.5 doubles it |
| API, Opus 5.5 | $3,318 | $39,818 | $16,591 | $199,092 | Each typical-to-heavy switch +$1,035 a month |
| Standard Compute | $1,560 | $18,720 | $7,800 | $93,600 | Fixed. Light on Economy ($39), typical on Standard ($89), heavy on Pro Plus ($499); only headcount and plan changes move it |
Before the reserve. Mix: 3 light, 5 typical and 2 heavy per 10 developers. Claude Team seats can't be sized in dollars because Anthropic doesn't publish the limits, so that row is a floor. The Copilot row assumes Copilot's agent uses tokens like our sample. Plan budgets: Economy $41, Standard $95, Pro Plus $549 a month.
What the example shows
Seat plans look cheapest and are the hardest to forecast for heavy agent users, because the cost past the limit is unknown until the month ends. The API is predictable per token and unpredictable in total. A fixed budget per developer costs about the same as the API for this mix and turns it into a number you can put in a spreadsheet; the trade-off is that someone has to resize a key that keeps hitting its budget.
Forecasting checklist for 2027
- Tier your developers from a two-week pilot, not a survey: give three people (your heaviest agent user and two typical ones) their own key and read their spend per active day.
- Plan for more autonomy. Parallel agents, background runs and longer sessions move developers from typical to heavy, and that step is the largest single change in the worksheet. See parallel agents in Claude Code.
- Decide the model mix. A day on Opus 5.5 costs twice a Sonnet 5.5 day, and a GPT-6 Luna day costs about 6% of one. Set a default model policy before the budget, not after.
- Put retirements and price changes on the calendar. Copilot retires Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra on November 16, 2026, and suggests Opus 5.5 (cheaper per token than Opus 5 at list price) or GPT-6.1 Sol and GPT-6 Luna instead; Anthropic retires Claude Sonnet 4.5 on November 30 and OpenAI the GPT-4o and GPT-4.1 families on January 14, 2027. A retirement can move cost either way. The running list: AI coding plan changes.
- Set the overage rule. Copilot's paid usage past the credit pool is on by default for organizations; decide who may approve extra usage on any plan, and set a budget or turn it off.
- Choose yearly or monthly billing knowingly. Yearly Claude Team and ChatGPT Business seats cost 20% less but fix headcount; Claude Enterprise needs 20 seats billed yearly.
- Check region and privacy needs early. Keeping inference in Europe or the US narrows the models and providers available, which can raise the cost of a day.
- Re-tier every month. Move developers between tiers or plans from real spend, so a heavy user gets a bigger budget before it blocks a sprint.
The catch
These figures come from list prices and our pooled agent days, not from your team. Your codebase, your agent and how long sessions run can move a typical day by several times, and the 3 / 5 / 2 mix is our assumption. A busy day every working day is an upper bound, so the heavy tier is generous on purpose. Run the pilot before you commit the number, and expect per-token prices and seat terms to change at least once during 2027.
Where Standard Compute fits
Teams that want AI coding spend lower and known in advance. The router sends each step to the cheapest capable model, with providers vetted for price, uptime and speed and automatic fallback. Opt-in Compliance rules keep inference in Europe or the US for the company, a team or a key, and we don't train on prompts. Each developer gets a key with its own plan (Business plans from $499, or pay as you go at usage plus 15%), invoiced monthly to the company. Creating the company is free.
FAQ
How much should a company budget per developer for AI coding agents in 2027?
At Claude Sonnet 5.5 list prices, about $37 a month for a light user, $73 for a typical one and $591 for a heavy one. With 3 light, 5 typical and 2 heavy per 10 developers that averages $166 a month, about $2,389 a year with a 20% reserve. Opus 5.5 doubles it; efficient models such as GPT-6 Luna cost a fraction.
Is a per-seat plan like Claude Team, ChatGPT Business or Copilot Business enough?
For light and typical users, often yes. Heavy agent users hit the 5-hour, weekly or credit limits and then pay extra usage at API rates, so their real cost is close to the API cost plus the seat. On Copilot Business the $19 seat is a $19 credit pool, and paid usage past it is on by default.
How do we stop one developer from using the whole team's budget?
Give each developer their own limit. Anthropic's console caps spend per organization or workspace, which stops everyone in it at once. Copilot lets admins set a budget or turn paid usage off. On Standard Compute every developer has their own key and plan, and a key at its budget stops with HTTP 402 while the others keep working.
Should we buy seats yearly or monthly for 2027?
Yearly is 20% cheaper on Claude Team ($20 vs $25 a Standard seat) and ChatGPT Business ($20 vs $25), but it fixes headcount and locks you to one vendor's models while plans and retirements change. One hedge: yearly seats for the stable core, and heavy agent use on a monthly, usage-based or per-key budget.
What could change our AI coding costs during 2027?
Three things: developers moving to more autonomous and parallel agents, the default model (Opus-class models cost about twice Sonnet-class ones), and vendor changes. In 2026 alone Copilot moved to AI credits on June 1 and retires three models on November 16. Keep a reserve on usage-billed lines and review tiers monthly.
What is the cheapest way to budget AI coding agents for a team?
Pay only frontier prices for the steps that need a frontier model. Routing routine steps to efficient models lowers the per-token bill, and a fixed budget per developer key keeps the total known. Measure two weeks of real use first, then size each developer from it.