AI coding agent pricing compared (2026): what 12 agents charge
AI coding agent pricing in 2026, side by side: what 12 vendors charge, what each one meters, and whether your own key or subscription counts in the cloud.
The plan price is the least useful number in AI coding agent pricing. Nine of the twelve products below have a tier at or near $20 a month, and those nine tiers buy nine different things: a seat, a pool of credits, a rate limit, a share of a prepaid balance, or $20 of usage at a marked-up rate. Two $20 plans can differ by a factor of ten once the agent starts reading a large repository. This post lines the twelve up on what actually varies, which is what the meter counts and whether a key or subscription you already pay for is allowed to do the work.
Prices and mechanics checked in early September 2026 against each vendor's own pricing and docs pages. Several of these changed in the last ninety days. The linked pages are the source; this table is a snapshot.
Sinatra is our product, so that row is a vendor describing itself. Every other row links to where we read it, and where the number is unpublished we say so rather than guess.
The table
| Agent | Paid entry | What the meter counts | Your key or subscription runs the cloud agent | Free tier |
|---|---|---|---|---|
| Sinatra | $20 per workspace member | Nothing on your own key; managed runs at the provider's exact rate | Yes: Anthropic or OpenRouter key, Claude or ChatGPT subscription | 5 tasks a day with your own API key or subscription, no card |
| GitHub Copilot coding agent | Pro $10/mo; Business $19 a seat | AI credits at $0.01 each, drawn per token at published rates, plus Actions minutes | No: BYOK covers editors and the CLI, not the cloud agent | Copilot Free excludes the agent |
| OpenAI Codex cloud | ChatGPT Plus $20/mo; Business $20 a user billed yearly | A rate limit per plan tier, not a rate | Only a ChatGPT plan; an API key turns cloud features off | ChatGPT Free, with limited usage |
| Google Jules | Google AI Pro, $19.99/mo | Tasks per day: 15, 100, or 300 depending on plan | No: Gemini only, no key | 15 tasks a day, 3 concurrent |
| Cursor Cloud Agents | Pro $20/mo; Teams $40 a seat | Model usage at the model's API price out of a usage pool; $0.25 per million tokens on top for Teams and Enterprise | Partial: custom keys documented for chat models, not Cloud Agents; Teams charges the token rate on your key too | Hobby is free but has no Cloud Agents |
| Claude Code (Action and web) | Nothing beyond your credential | Your Anthropic bill at token rates, or your Claude subscription; Actions minutes on GitHub | Yes: API key, Pro/Max/Team/Enterprise subscription, or Bedrock/Vertex/Foundry | No managed tier, so no free tier |
| Linear coding sessions | Linear Basic $10 a user; Business $16 | Prepaid dollars: model tokens at provider-published rates, no markup, plus $0.25 per 20 minutes of sandbox | No: a workspace-selected model, billed from the credit balance | None; you top up first |
| Devin | Pro $20/mo; Teams $80/mo plus $40 a seat | An included daily and weekly allowance of unpublished size, then on-demand credits; Enterprise is metered in ACUs at an order-form rate | No key path documented | A free tier with an unpublished allowance |
| OpenHands | $0 to self-host; Enterprise custom | Model usage at cost with no markup on the hosted tier | Yes for keys; subscriptions only where the CLI login already sits on the machine | 10 hosted conversations a day, unlimited locally |
| Factory | Pro $20/mo; Teams $60/mo plus $40 a seat | Rolling 5-hour, 7-day, and 30-day rate limits plus prepaid credits | Partial: CLI and desktop only, and free only up to an allowance | None |
| Kiro | Pro $20 a user | Credits: 1,000 on Pro, add-ons $0.04 each, no published token conversion | No | 50 credits, once |
| Augment Code | Standard $20/mo for up to 50 seats | $20 of usage, with a flat 40% fee on model usage | No | None |
The four shapes AI coding agent pricing takes
The twelve sort into four billing shapes, and the shape tells you more than the price.
Your provider's invoice, plus a fee for the machinery
Sinatra, Claude Code, and OpenHands all leave the model bill with your model provider. You hold the key or the subscription, the tokens land on that account at whatever rate you already have, and the vendor charges for something else or for nothing.
Claude Code charges nothing at all. You supply the credential, Anthropic bills the tokens, and GitHub bills the Actions minutes if you run it as an Action. There is no managed tier, which is why it has no free tier either: someone always has to bring a key.
OpenHands is free to run on your own hardware and free on its hosted Individual plan up to 10 conversations a day. Where it bills models on your behalf, its pricing page says at cost with no markup. That is the cleanest price list in this market and we would rather say so than pretend otherwise.
Sinatra charges $20 per workspace member, and adds nothing to the model side at any volume. On your own Anthropic or OpenRouter key the tokens are on your invoice. On a Claude or ChatGPT subscription the marginal cost of a task is whatever that subscription already costs you. Leave the credential unset and a paid workspace runs on our keys at the provider's exact rate from a prepaid balance.
Published per-token rates, billed by the vendor
Copilot, Cursor, and Linear publish a rate and bill you for it themselves. You can see what a session cost after it runs, at a price you could have looked up beforehand. You cannot bring a relationship you already have with a model provider, though; the vendor holds that one.
GitHub moved Copilot to usage-based billing in June 2026. Each plan includes a pot of AI credits at a cent each (1,000 base credits a month on Pro, 1,900 per user on Business), and the credits drain per token at published per-model rates that match Anthropic's own list prices row for row on the day we checked. The agent also burns Actions minutes, because it runs inside a GitHub Actions environment with a 59-minute cap per session. Copilot's bring-your-own-key covers editors, the CLI, and the SDK; the agent on GitHub.com is not on that list.
Cursor is more transparent than most. Third-party models are charged at the model's API price, and its docs put a number on it: daily agent users typically land at $60 to $100 a month of total usage. The catch sits on the team plans. Teams and Enterprise add a Cursor Token Rate of $0.25 per million tokens on top of model API pricing, and the docs say that applies to BYOK usage as well. Small and published, but still a per-token fee on a key you bought yourself.
Linear coding sessions draw on a prepaid balance in dollars, with model tokens at provider-published rates with no markup and sandbox runtime at $0.25 per 20-minute block. That is the only pricing on this list where you can estimate a session before running it. There is no free allowance and no way to point it at a subscription you hold; you top up and the balance does the paying.
A rate limit instead of a rate
Codex, Factory, and Devin sell tiers, and what a tier buys is throughput.
Codex cloud comes with a ChatGPT plan. OpenAI publishes an estimate rather than a price: roughly 10 to 100 local messages per five hours on Plus for the model cloud tasks run on, with the caveats that cloud chats may use more of the allowance and that the estimates are not fixed limits. We went through the whole plan table in a separate post on Codex cloud pricing. When Plus is not enough, the answer is Pro at $100 rather than a bill that tracks usage. There is a per-token path through an API key, and OpenAI's pricing page says that path has no cloud features. You get published pricing or the cloud agent, not both.
Factory's plans run $20, $100, and $200 for individuals, or $60 a month plus $40 a seat for Teams, against three rolling rate-limit windows of unpublished size. Its own docs say bring-your-own-key is free up to an allowance, after which usage is charged according to your plan. A key lowers the bill without decoupling it from usage, which is the thing a key is usually brought to do.
Devin's Pro plan is $20 and its Teams plan is $80 plus $40 per full seat. Cognition's self-serve billing page says Pro and Teams seats have a daily and weekly allowance, Max has a weekly one, and usage past that comes from on-demand credits. None of the allowances is given as a number. Two tickets that look identical when written can cost different amounts, and you learn which afterwards.
Jules belongs here too, though it meters something simpler: tasks. Fifteen a day free, 100 on Google AI Pro, 300 on Ultra, with no separate model bill. For an individual on GitHub who is happy with Gemini, that is generous and easy to reason about.
A unit the vendor defines
Kiro's credits, Devin's Enterprise ACUs, and Augment's fee are the shape to read most carefully, because the unit is the product.
Kiro publishes plenty: 50 free credits, 1,000 on Pro at $20, up to 10,000 on Power at $200, and add-ons at $0.04 each. Its pricing page says a simple prompt can use less than a credit and a spec task usually more, and its models page gives per-model multipliers, but nowhere does Kiro say what a credit is in tokens. You find out how big the unit is by spending it.
Augment publishes something few others do: a flat 40% fee on model usage, with no fee on compute. The $20 Standard plan buys $20 of usage for up to 50 seats, which at low volume may be the cheapest thing on this page. The percentage is the part to sit with, because it scales with exactly the thing you are trying to do more of. A team that doubles its agent usage doubles the fee.
The number almost nobody publishes
The number you actually want is the cost of one task, and only Linear lets you estimate it before you run it. Copilot and Cursor tell you afterwards at a rate you could have looked up. Codex and Factory tell you when you hit a window. Kiro and Devin tell you in a unit you cannot convert.
The way around that is to stop reading pricing pages, including this one. Take five tickets you would actually hand to an agent, run the same five through each product on your shortlist, and read the invoices next to the pull requests. Where the model bill lands on your own provider account, count that too. An afternoon of that beats any table, and most of these vendors have a free tier or a trial that covers it.
How Sinatra prices it
Two numbers, both published. The seat is $20 per workspace member, flat, and it does not move when a week gets busy. The model bill is your provider's, at their rate, or nothing at all if the credential you connect is a Claude or ChatGPT subscription you already hold. For a small team the cheapest honest path is usually the subscriptions they already pay for plus $20 a member for orchestration. The free tier is 5 tasks a day with your own API key or subscription, with no card. We wrote up the tradeoffs of holding your own key separately, and the alternatives hub has a page against each product above.
Common questions
Which AI coding agent is cheapest?
For one person on GitHub, Jules: 15 free tasks a day with no model bill. For a team that wants to own the infrastructure, OpenHands, which costs nothing to run yourself. For a team whose developers already pay for Claude or ChatGPT, an agent that runs on those subscriptions, which today means Codex cloud for ChatGPT plans, Claude Code for Claude plans, or Sinatra for either. Everything else is cheapest at some volume and expensive at another, and the crossover depends on how many tokens your repository makes an agent read.
Does "bring your own key" mean no markup?
Not on its own. Cursor takes your key and, on Teams, charges $0.25 per million tokens against it. Factory takes your key in the CLI and meters it past an allowance. Copilot takes your key in the editor and not in the cloud agent. The question to ask is narrower: does my key run the hosted agent, and does anything get added to what my provider charges? Claude Code, OpenHands, and Sinatra answer yes and no.
What is an ACU?
Cognition's unit for Devin's work on Enterprise plans, priced in each customer's order form. Self-serve plans no longer meter in ACUs: they come with an included allowance and then on-demand credits, which the docs say carry the same dollar value an ACU did. Either way the unit bundles the agent's actions and machine time into one number. That is reasonable for billing compute and nearly useless for guessing what a ticket will cost before you run it.
Is a free tier enough to evaluate on?
Sometimes. Jules gives you 15 tasks a day, OpenHands 10 hosted conversations, and Sinatra 5 tasks, all on the same agent paying customers get. Copilot Free and Cursor Hobby exclude the cloud agent entirely, so there is no free way to try either on a real ticket. Augment, Factory, and Linear coding sessions have no free allowance at all. Devin and Kiro have free tiers whose size is either unpublished or spent once.
If you want to run the five-ticket test on Sinatra, start for free and the model bill stays on whatever you already pay for. The pricing page has the rest of the terms.