AI coding assistant pricing: what teams actually pay in 2026

AI coding assistants cost between $0 and $200 per developer per month on subscription, and the paid entry point for a capable tool is around $10 to $20. GitHub Copilot Pro is $10 per user per month, Cursor and Claude Code are both $20, Devin Pro is $20, and Agentcode starts at $29 a month billed yearly. The number that catches teams out is not the subscription. It is that most of these products meter agent usage on top of it, so the plan price is a floor rather than a ceiling.

This guide lists what each vendor publishes as of July 2026, explains where the metered costs hide, and gives you a way to work out your real monthly bill before you commit a team to it.

AI coding assistant pricing compared

All figures below are as published by each vendor and checked in July 2026. Prices in this category move quickly, so confirm on the vendor's own pricing page before you buy.

One correction worth making, because most articles on this topic still get it wrong: Devin's old $500 a month team plan was retired in April 2026, and its self-serve plans no longer bill in ACUs. If you read a comparison quoting $500 a month or $2.25 per ACU, it is out of date.

How much does GitHub Copilot cost?

GitHub Copilot costs $10 per user per month on Pro, $39 on Pro+, and $100 on Max, with organization plans at $19 per seat for Business and $39 per seat for Enterprise. There is a free tier with roughly 2,000 completions a month. All paid plans include the cloud agent, which can take a GitHub issue and open a pull request.

Copilot is the cheapest credible entry into agentic coding, and that is a genuine advantage rather than a loss leader. The limits worth pricing in are structural rather than financial: the cloud agent runs on GitHub-hosted repositories only, caps each session at 59 minutes, and works on one repository and one branch at a time. If your code is on GitLab, the cheapest plan on the market is not available to you at all. We go through that in detail on the GitHub Copilot alternative page.

How much does Cursor cost?

Cursor is free on the Hobby tier, $20 a month for the individual plan, $60 for Pro Plus, $200 for Ultra, and from $40 per user per month for Teams. The subscription includes a dollar-denominated pool of API usage, which is the detail that matters.

Cursor's cloud agents bill at the underlying model's API pricing, drawn against that included pool. Once the pool is spent, you either move up a tier or pay as you go against a spend limit you set. For a developer who mostly uses Tab completion and inline edits, the $20 plan is comfortably enough. For a team leaning on cloud agents for real work, the effective cost is the subscription plus whatever the agent runs consume, and that second number is the one to model. See the Cursor alternative page for the full picture.

How much does Claude Code cost?

Claude Code is included with Claude Pro at $20 a month, or $17 a month billed annually. Max plans start at $100 a month for 5x usage and $200 for 20x. Team standard seats are $25 a month, or $20 billed annually, and premium seats are $125.

Anthropic does not publish numeric usage limits, only relative multipliers, which makes capacity planning genuinely awkward: you cannot compute in advance how many tasks a 5x plan buys you. What Anthropic does publish, usefully, is an average API cost of roughly $13 per developer per active day, with most users staying under $30 per active day. That figure is a better basis for a budget than the subscription price, and it is a rare piece of honest disclosure in this category. Anthropic also notes that running tasks in parallel consumes your account rate limits proportionately, so a team hammering it concurrently will feel the ceiling sooner than the headline suggests.

How much does Devin cost?

Devin has a free tier, Pro at $20 a month, Max at $200 a month, and Teams at $80 a month plus $40 per full developer seat. Enterprise is custom, and is still billed in Agent Compute Units at a rate negotiated in the order form rather than published.

Devin's pricing has come a long way down since launch, and the current entry point is reasonable. The forecasting problem is that Cognition does not publish the size of the usage quota on each tier, describing it as an allowance that refreshes daily and weekly. Overage is billed as prepaid credits, which do at least roll over and never expire. A three-developer team on Teams starts at $200 a month before anyone runs a single task. The Devin alternative page compares the two models directly.

Where the real cost hides: metered agent usage

Every product in this table except Agentcode meters agent work in some form, whether it is called credits, an API pool, ACUs, or a usage multiplier. The reasoning is sound. An agent run that reads a repository, edits files across it, and executes a full test suite consumes far more compute than an autocomplete suggestion, and someone has to pay for it.

The consequence for a buyer is that per-seat sticker prices are not comparable across vendors. A $10 Copilot seat and a $20 Cursor seat are only equivalent for a developer who barely touches the agent. Once the agent is doing real work, the metered layer dominates, and it is the part you cannot see on the pricing page.

This is a familiar shape if you have watched cloud bills grow: consumption pricing is efficient right up until the month it surprises you. Teams that already watch what their cloud and SaaS spend does over time tend to catch this early, because the pattern is the same one they have seen with usage-based infrastructure. Teams that treat AI tooling as a fixed per-seat line item tend to find out in arrears.

How to work out your real monthly bill

You cannot get this from a pricing page, so run a two-week measurement instead. It is worth the effort, because the gap between the sticker price and the real one is routinely two or three times.

1. Count the tasks, not the seats. How many agent tasks a week will your team actually delegate? For most teams starting out it is far fewer than they predict, often three to five per developer per week.
2. Run a pilot on one repo. Give the agent real backlog work for two weeks on a paid plan and watch the credit or usage meter, not the clock.
3. Extrapolate the metered layer. Take the usage consumed in that fortnight, double it, and add it to the subscription. That is your realistic monthly number.
4. Price the failure rate too. An agent PR you reject still consumed usage. Track the ratio of merged to abandoned pull requests, because a tool with a 50% rejection rate costs twice what its meter suggests.
5. Compare against the alternative honestly. The benchmark is not zero. It is what an engineer's time costs to do the same task, and against a fully loaded US engineering salary the tooling is close to a rounding error if it works.

Is a free AI coding assistant good enough?

For evaluation, yes. For sustained work on a real codebase, no, and the vendors are open about why. Free tiers are sized for you to form an opinion, not to clear a backlog: Copilot's free tier is roughly 2,000 completions a month, Cursor's Hobby tier is explicitly limited, and Claude Code is not on the free Claude plan at all.

The economics are unavoidable. A single agent run that clones a repository, reads across it, edits multiple files, and executes a test suite costs the provider real money in inference and compute. Nobody gives that away at volume. If a tool's primary pitch is that it is free, the constraint has been moved somewhere you have not looked yet, usually into a quota that runs out in the first week.

What Agentcode costs, and why it is flat

Agentcode starts at $29 a month billed yearly for Starter, $59 for Pro, and $129 for Team, with a custom Enterprise tier. There is no free plan and no usage meter on top. You are not cheaper than Copilot Pro on sticker price, and we would rather say that plainly than bury it.

The flat model is a deliberate trade. You give up the theoretical savings of a light month, and in exchange you get a number your finance team can put in a spreadsheet and forget about. For engineering managers who have to defend a tooling budget, a predictable line item is often worth more than a lower expected value with a long tail. Full details are on the pricing page.

Which one should you buy?

Pick on the shape of the work rather than the price, because the price differences are small next to the cost of choosing the wrong category.

• You want faster typing. Buy Copilot at $10 or Cursor at $20. An in-editor assistant is what you want, and both are excellent at it.
• You want to delegate whole tasks and you are on GitHub. Copilot's cloud agent is included at $10 and is the cheapest way to try the agentic model. Watch the 59-minute session cap.
• You want to delegate whole tasks and you are on GitLab. Your options narrow immediately. Copilot's cloud agent and Claude Code's cloud sessions both require GitHub to open a pull request. Agentcode works on both.
• You want an agent fleet you drive from Slack or an API. Devin is built for that and the others are not.
• You need a predictable bill and a security review you can pass. That is the enterprise AI coding assistant conversation, where flat pricing, SSO, audit logs and a no-training guarantee matter more than $10 of monthly spread.

For the wider landscape of what these tools do rather than what they cost, read the best AI coding tools guide and the background on what an AI coding agent is.