AI coding agent cost for a 10-person startup, in real numbers

An AI coding agent for a startup costs two things, and the pricing pages only show you one. There is the seat, or the plan, or the credit pack, which is the number on the page. And there is the model bill, which is what the agent spends reading your repository and writing code, and which on most tools is either bundled into a unit you cannot inspect or billed by the vendor at a rate they set. For a ten-person team the second number is usually larger than the first, and whether you already pay for Claude or ChatGPT can change it by a factor of five. This post does the whole-number math for a small team across the tools a startup actually shortlists, with every assumption written down so you can swap in your own.

Competitor prices were checked in early September 2026 against each vendor's own pages, which are linked, and re-checked in October 2026, by which point Cursor's had changed. Sinatra's terms reflect its current plans and own-credential requirement. The token estimates are ours and are labeled as such.

Sinatra is our product. It is in the examples, the assumptions are the same for every tool, and the one where we come out worst is left in.

Ten engineers. Four of them hand tickets to an agent; the other six review pull requests, write code in an editor, and never trigger a run. Between them, the four delegate about forty tasks a month, two a day on a working week.

Each task is a real ticket on a real repository, so the agent reads a lot and writes a little. Our estimate, and it is only that: about 1.5 million input tokens and 100,000 output tokens per task on a frontier model. At Anthropic's list rates for Claude Opus 5 ($5 per million input tokens, $25 per million output), that is roughly $10 a task and $400 a month in tokens, before prompt caching brings it down. If your repository is small and your tickets are tight, halve it; if you point the agent at a monorepo with a slow install, double it. The point of the exercise is the shape of the bill, not the third decimal.

One more thing to hold onto. On several of these tools the seat is only owed for people who use the agent, so the ten-person team is often a four-seat bill.

Copilot Business is $19 a seat a month, and since June 2026 each seat includes 1,900 AI credits a month, which at a cent each is $19 of model usage at GitHub's published per-token rates. Those rates matched Anthropic's list prices row for row when we checked. The agent also runs in a GitHub Actions environment, so its sessions consume Actions minutes.

Four seats is $76, and it includes $76 of credits. Forty tasks at our estimate is $400 of tokens, so about $324 of that is billed as additional usage on top. Call it $400 a month plus Actions minutes for the four-seat case. If all ten engineers hold a Copilot seat for the editor anyway, the bill is $400 to $514: credits are pooled across the organization, so whatever the other six leave unused offsets the agent's overage, and you were paying for those seats already.

What does not help: the Claude Max or ChatGPT plan on anyone's card. Copilot's bring-your-own-key covers editors and the CLI, not the cloud agent, so every token goes through GitHub's billing.

Cursor is now sold as part of The X Subscription, and a Teams seat is $30, $100, or $200 a month, with usage that matches the individual plan at the same price. Premium, at $30, is the first tier with Cloud Agents. Cursor's own models draw on the seat's usage; Claude, GPT, and other third-party models are billed at their API list price, and on Teams a Cursor Token Rate of $0.25 per million tokens comes on top, including on tokens bought with your own key (pricing docs).

Four $30 seats is $120. Forty tasks on a third-party model at API price is about $400, and the token rate on 64 million tokens adds about $16. So roughly $536 a month if only the four delegators are on Cursor, and about $416 of new spend if all ten already pay for Cursor as their editor. That editor is the reason a lot of startups will land here regardless of the arithmetic, and it is a good reason.

Subscriptions do not count. Cursor's usage pool cannot draw on a Claude or ChatGPT plan.

Devin Teams is $80 a month plus $40 per full seat on Cognition's pricing page, which makes four seats $240 before any usage. Its self-serve billing page reads differently: an $80 monthly minimum that covers two full seats, then $40 a seat, which would make four seats $160. Ask before you budget. Either way, Pro and Teams seats come with a daily and weekly allowance, and usage past it draws on on-demand credits. The size of the allowance is not published.

That makes this the one example we cannot finish. $160 or $240 is the floor. Whether forty tasks fit inside the allowance, and what the overage costs when they do not, is something you learn by running them. No key path is documented. A linked ChatGPT Go, Plus, or Pro plan can cover Devin's calls to OpenAI models, but the rest still bills to the Devin quota, so a subscription changes the number only at the edges.

If the team already uses Linear on Basic or Business, the agent has no seat cost at all. Sessions draw on a prepaid balance at model tokens at provider-published rates with no markup, plus $0.25 per 20-minute block of sandbox. At our estimate, forty sessions of about forty minutes each is $400 of tokens and $20 of sandbox time, so about $420 a month, and you can work that out before you run anything, which is rare.

Subscriptions do not count here either; the balance does the paying. And there is no GitHub-issue trigger, which matters for a startup with repositories whose issues never moved into Linear. We went through the whole thing in Linear coding sessions, explained.

Sinatra is a flat price per workspace, not per person. A workspace connects at most one GitHub organization or personal account, and owners, admins, delegators, and members who only review all sit under the one price. What the plan sets is how many agents work at once: Pro is $20 a month for 3 agents at a time, plus 10 waiting; Team is $99 for 10 agents at a time, plus 25 waiting. Forty tasks a month from four people fits Pro; a ticket assigned while three are already running waits and starts on its own when one finishes. Whether the other six join the workspace does not change the bill. The examples below assume Pro.

The model bill depends on what the four connect. On an Anthropic or OpenRouter key, the tokens are on your provider's invoice at their rate, so the same forty tasks cost about $400 and the month is about $420. That is the same band as Copilot and Cursor, because at API rates tokens dominate everything and nobody on this page is marking them up by much.

The number changes when the four already pay for Claude or ChatGPT. Each member can connect their own subscription, tasks they start run on it, and the workspace needs no shared key. The marginal cost of a task is then whatever the subscription already costs, inside its own limits, and the month is $20. That is the whole cost: the subscriptions you already pay for, plus $20 a month for the whole workspace.

Each personal subscription is available only to the member who connected it. Sinatra tasks share that person's provider allowance with their other usage, and provider rate limits still apply. Paid workspaces also require a personal connection or workspace API key; leaving credentials unset does not provide model access. Token and cost diagnostics are estimates, not amounts owed to Sinatra, including when a subscription covers usage.

For teams under about a hundred people, Team is half price for a year. The startup program takes 50% off Team for 12 months: $49.50 a month for 10 agents at a time, plus 25 waiting, for the whole workspace. The discount is on Team only, so for the team in this post it is not the cheapest option: forty tasks a month fits Pro at $20. The program is worth applying for once the four delegators routinely want more than three tickets running at once. The model side is unchanged either way: your provider bills you directly at their rates, or your subscriptions cover it, and the discount is on the Sinatra plan only. You apply with a company URL and a team size, a person reads it, and if it fits you get a single-use code to enter at checkout. Checkout still collects a card, billing stays month to month, and at month 13 Team returns to $99.

There is no funding stage to prove and no investor list. The bucket above 100 is on the form on purpose, because we would rather read the application than turn a team away at the door.

Tool	Seats or plan for the four delegators	Model bill at API rates	With Claude or ChatGPT subscriptions	Month, four delegators
Copilot coding agent	$76	$400, of which $76 is included	No change; subscriptions do not apply	About $400 plus Actions minutes
Cursor Cloud Agents	$120	$400 plus $16 token rate	No change	About $536
Devin Teams	$160 to $240	Inside an unpublished allowance, then credits	A linked ChatGPT plan covers only its OpenAI calls	$160 to $240 floor, rest unknown
Linear coding sessions	$0 on an existing Linear plan	$400 plus $20 sandbox	No change	About $420
Sinatra	$20 for Pro, whole workspace, any number of people (Team $99, or $49.50 on the startup program)	$400 on your own key	$0 additional within existing subscription limits	$420 on API keys, or $20 on existing subscriptions, excluding subscriptions already paid for

Read the table as a shape rather than a ranking. At API rates everyone lands between $400 and $600, and the choice is about triggers, editors, and who you want to hold the model relationship. The gap opens only when the people delegating already have subscriptions, and then it is $20 against $400 and up.

Usage grows faster than headcount. The four delegators become eight, the two tasks a day become five, and anything metered per token or per credit grows with it. A per-seat price grows with the people, a percentage fee grows with the usage most of all, and a flat workspace plan grows with neither: more people cost nothing, and more volume means tasks spend longer waiting until you decide you want more running at once, which is a step from $20 to $99 rather than a meter. Augment's published 40% on model usage is the extreme case, and we covered it in the pricing roundup.

Units hide the growth. A credit, an ACU, or a plan tier tells you that you ran out; a provider invoice tells you why. Prefer the tool that lets you see the token bill even if you decide to let the vendor pay it.

And the rollout path matters at ten people because it will matter more at forty. On Sinatra, a workspace connects one GitHub organization, invited members bring their own subscription or use the workspace key, and each repository can carry a committed .sinatrarc for its model and test command. The BYOK post covers what holding the key yourself asks of you.

If you want to run the forty tasks before deciding, start free with the subscriptions your team already has, and apply to the startup program when the free tier stops being enough.