Devin's pricing page says $20 a month. Reviewers who use it report bills of $300–500. Both numbers are real, and the distance between them is the ACU — Devin's usage meter. As of July 2026, Devin Pro is $20/month, Max is $200/month, and Teams is $80/month plus $40 per seat, and on top of every plan sits metered usage: historically billed as Agent Compute Units at about $2.25 each, one ACU covering roughly 15 minutes of active agent work.
This post walks through the meter, runs the math on a realistic month, and shows where the model is a good deal and where it isn't. We build Codowave, a competing agent with flat-quota pricing, so we have a position here — but the arithmetic below is arithmetic, and you can re-run all of it yourself.
How Devin pricing works in 2026
Devin's pricing has been through three eras, and articles from each era are still floating around, which is why search results contradict each other.
Era one was the launch price: $500/month, no self-serve entry. Era two arrived in April 2025, when Devin 2.0 cut the entry price to $20/month and moved to pay-as-you-go ACUs at roughly $2.25 each on the Core plan; the legacy $500 Team plan bundled 250 ACUs at a discounted $2.00 (Costbench's 2026 breakdown). Era three is the current page. As of July 2026, devin.ai/pricing lists:
- Free — a light daily quota for trying agents.
- Pro — $20/month, higher quotas, full model access, Devin Cloud.
- Max — $200/month, everything in Pro with much higher quotas.
- Teams — $80/month plus $40 per seat, pooled usage and billing.
- Enterprise — custom, with SSO and VPC deployment.
The structural constant across all three eras: the subscription buys access and an allowance; the work is metered. Current plans refresh usage allowances daily and weekly, and overage is billed "at API pricing" — an open-ended term worth reading twice if you're the one approving the budget.
What an ACU is and what one costs
An ACU — Agent Compute Unit — is Cognition's normalized unit for the resources Devin consumes while actively working: VM time, model inference, and bandwidth. One ACU corresponds to roughly 15 minutes of active agent work, and it ran about $2.25 on the self-serve plan (eesel's pricing analysis, 2026). Even as the current plans reframe usage as daily and weekly allowances, the ACU remains the unit underneath, and it's the right mental model for estimating cost.
The property that matters most is what the meter measures. An ACU meters effort, not outcomes. A session that ends in a merged PR and a session that ends in a PR you close unread consume ACUs at the same rate. That's not a criticism of Cognition specifically — every usage-billed agent works this way — but it defines when the math works.
Per-task consumption is the number to estimate. Reviewers in 2026 put small, well-scoped tasks at roughly 2–5 ACUs and multi-file features at 10–30, with runaway sessions going higher (Brainroad's cost teardown, 2026). Cognition's own guidance has always pointed the same direction: keep sessions short and tightly scoped.
A worked example: one team, one month
Take a four-person team that routes 40 tickets a month through Devin, with a distribution any engineering lead will recognize:
- 25 small fixes at ~3 ACUs each = 75 ACUs
- 12 medium features at ~10 ACUs each = 120 ACUs
- 3 tasks that go sideways at ~25 ACUs each = 75 ACUs
That's 270 ACUs. At the $2.25 self-serve rate, the usage alone is about $608. On Teams, the base is $80 plus four seats at $40 — $240 — so the month lands near $850 before taxes. If 34 of the 40 tickets produce merged PRs, that's roughly $25 per merged PR.
Now the small-scale version: a solo developer delegating 10 well-scoped tickets at ~3 ACUs each burns about 30 ACUs — roughly $68 of usage on a $20 plan, some of it inside the included allowance. Call it under $100 total.
Both scenarios are consistent with what independent reviewers report: light use stays near the sticker price, active use lands at $300–500 a month. Neither number is hidden; they're just not the number on the front of the pricing page.
When the ACU math works
At $25 per merged PR, Devin is cheap — that's the honest headline. If each of those tickets would have cost an hour of a $75–150/hour engineer, the team in the example bought back thousands of dollars of time for $850. The math works best on exactly the work Devin is best at: migrations and repetitive pattern tasks, where Cognition's own 2025 performance data shows file migrations finishing in 3–4 hours against 30–40 human hours. High ACU burn, higher hours saved.
It also works at low volume. A developer who delegates a few tasks a month never leaves the allowance, and $20/month for an occasional autonomous teammate is a fine trade.
When the math breaks down
Three situations turn the meter against you.
Vague tickets. An agent handed "improve performance" loops, retries, and reads more files — and every minute of confusion bills at the same rate as a minute of progress. Effort-metered billing charges failures at full price.
Volume. The meter scales linearly with your backlog. The 40-ticket team pays ~$850; the same team at 100 tickets is modeling a $2,000 month. Per-PR economics stay fine, but the absolute number keeps climbing with usage, and there's no ceiling built into the plan.
Budgeting. Usage varies month to month, and overage "at API pricing" is open-ended by construction. If your approval process wants a fixed number, a usage-billed agent makes you forecast your own backlog's difficulty — which is the thing nobody can forecast. We've written before about why hard cost ceilings are the feature that lets finance say yes.
A usage meter bills the agent's effort. You get paid for the agent's outcomes. The gap between those two curves is your risk.
Devin cost comparison vs alternatives
Sources: devin.ai/pricing (July 2026), codowave.com/pricing, our 2026 agent field guide.
| Tool | Pricing model | Entry price | The bill scales with | Worst-case month |
|---|---|---|---|---|
| Devin Pro | Subscription + metered usage | $20/mo | Agent minutes (ACUs) | Open-ended — overage at API pricing |
| Devin Teams | Base + seats + pooled usage | $80/mo + $40/seat | Agent minutes across the team | Open-ended |
| Codowave | Flat plan with hard issue quota | $19/mo (500 issues) or $49/mo (1,500) | Nothing — quota is a hard cap, no overage | Plan price + the BYOK LLM cap you set |
| OpenAI Codex | Usage-based token credits | Usage | Tokens consumed | Typically low hundreds per dev |
| Cursor | Flat subscription + request limits | $20/mo | Requests past quota | Modest |
The Codowave row deserves its own honesty check, since it's ours. The flat price covers the managed infrastructure and a hard monthly quota of tracker issues — when the quota's spent, new work waits for the next cycle rather than billing overage. The LLM itself is bring-your-own-key Anthropic Claude, billed by Anthropic directly under a monthly spend cap you configure. So "flat" means the Codowave line on your budget is flat and the LLM line has a ceiling you chose — not that inference is free. And the fit boundary is real: Codowave works your issue tracker; it isn't an IDE pair-programmer or a prompt-driven one-off task runner, which is exactly where Devin's model earns its keep.
The one-line summary of Devin pricing: $20 is the door, the ACU is the meal. Decide whether it's worth it by estimating the meal — 270 ACUs for a 40-ticket month — and never by the number on the door.