When Cognition AI unveiled Devin in March 2024, the industry's reaction was mixed. Some hailed it as the "first AI engineer," while others questioned its real-world utility. However, following its public launch and the reveal of its pricing structure, it became clear that Devin is positioned not as a hobbyist tool, but as an enterprise-grade product with a corresponding price tag.

For AI builders considering Devin for their team or personal projects, it's crucial to grasp not just the initial cost, but also the billing logic centered around ACU (Agent Compute Units). This metric ultimately dictates the true cost of each session.

This article serves as a practical guide: what's included in the free plan, where the paid tiers begin, and what limitations to be aware of before your first integration.

What are ACUs and how Devin tracks usage

Devin doesn't bill per month or per session count. Instead, it uses ACU – Agent Compute Units. This is an internal metric that reflects the computational expense of task execution: how many steps the agent took, how many tools it invoked, and the complexity of the context it handled.

One ACU roughly equates to several minutes of active agent work. A simple task, like fixing a bug in a function, might cost 1–3 ACUs. A complex refactoring task involving tests and opening a Pull Request could range from 10–30 ACUs or more.

It's important to note that ACUs are not a one-to-one conversion with tokens. Cognition deliberately abstracted away from tokens to simplify cost planning for teams. This is convenient for managers but makes precise cost prediction for non-standard tasks more challenging.

Pricing tiers: Free, Team, and Enterprise

As of 2025, Devin offers three access levels:

• Free Plan. Includes trial ACUs upon registration, sufficient for a few exploratory sessions. It limits users to one concurrent session and offers basic documentation as the sole support channel. This plan is suitable for evaluating capabilities, not for productive work.
• Team (~$500 USD/month). Provides a fixed monthly ACU package, allows for multiple parallel agents, integrates with GitHub, Slack, and Jira, and offers priority support. This is an optimal option for smaller teams delegating routine coding tasks.
• Enterprise (By request). Offers a custom ACU package, Service Level Agreements (SLAs), Single Sign-On (SSO), audit logs, and the option for private deployment with a dedicated account manager. This tier is designed for companies with stringent security and reproducibility requirements.

Additional ACUs can be purchased on top of the base package. The per-unit cost decreases as the volume increases. It's always advisable to check the official Cognition page for the most current figures, as the pricing structure has been adjusted since the public launch.

The real cost of a session: What influences ACU expenditure

The theoretical limits and the practical cost of a task can differ significantly. Here are the key factors:

• Task Clarity. Devin performs better with structured instructions (e.g., "write a unit test for this function covering these edge cases") than with open-ended requests (e.g., "improve the project's architecture"). Vague tasks lead to unnecessary iterations, thus increasing ACU costs.
• Codebase Quality. A well-prepared repository—with clear READMEs, comments, and clean code free of hidden dependencies—allows the agent to spend less time on reconnaissance. A neglected codebase can consume ACUs before any meaningful progress is made.
• Environment Readiness. If Devin has to set up the development environment from scratch—fixing `npm install` issues, finding missing environment variables, or resolving dependency conflicts—this directly impacts the ACU count. A stable CI environment pays for itself.
• Parallel Agents. On the Team plan, multiple agents can run concurrently. This is useful for scaling operations, but ACUs are consumed in parallel, depleting the balance faster.

A practical tip: before embarking on a large undertaking, break it down into sub-tasks and assign them to Devin one by one. This makes it easier to control costs and verify results at each stage, preventing you from exhausting your entire monthly limit on a single unsuccessful prompt.

AiiN's Conclusion

Devin is not like Claude or GPT for code, which answer questions in a chat interface. It's an autonomous agent to which you delegate an entire task and await a Pull Request. Its pricing threshold of approximately $500/month makes it less relevant for solo developers or student projects, but justifiable for small teams where the agent can handle a portion of routine coding and code review tasks.

The free trial is sufficient to determine if the tool fits your workflow. If, after a few sessions, ACU costs have accumulated without tangible results, it's worth re-evaluating your task formulations and repository preparation before immediately upgrading to a paid plan.

For AI builders already using tools like Cursor or GitHub Copilot, Devin addresses a different segment: not real-time code completion, but the delegation of entire tasks with minimal oversight. Both approaches can coexist within a single tech stack, depending on which parts of the workload you're willing to fully hand over to an agent.