Devin's $20 Plan Sounds Cheap. Here's What You Actually Pay.

Cognition dropped Devin’s headline price 96%—from $500 to $20/month—and the internet celebrated. Then people started invoicing themselves and realized the math was a lie. Devin’s official pricing gate-keeps usage behind Agent Compute Units (ACUs), and the $20 gets you about 9 of them. After that, you’re paying $2.25 per ACU whether the task succeeds or fails.

We’ve spent the last month watching Devin invoices land in dev Slacks across three companies. None of them were $20.

The Headline Number Is $20. The Real Number Is Closer To Your Replit Bill.

Devin Docs confirm the structure: the Pro plan at $20/month includes a quota of ~9 ACUs. Each additional ACU costs $2.25. That’s the entire trap. The plan is real. The price is real. The monthly bill is not.

One developer working a single 15-minute troubleshooting session burns about 1 ACU. Two sessions a day, five days a week—that’s 40 ACUs/month on top of the included 9. You’re at roughly $70/month before you’ve shipped anything. Add vague prompts, re-runs on failed tasks, or a codebase the agent needs to navigate—suddenly that bill doubles or triples.

BrainRoad’s real-cost breakdown estimates a light solo dev at ~$70/month and a small engineering team running 200-300 ACUs at ~$520/month. Neither is the marketing number.

What Is an ACU and Why the Definition Matters

An ACU (Agent Compute Unit) is Cognition’s metering unit. One ACU ≈ 15 minutes of Devin’s active work—VM uptime, inference tokens, network calls, the whole thing. According to the billing docs, the meter runs on activity, not output. A failed run still burns ACUs. A clarification loop burns ACUs. A re-prompt because your initial spec was vague burns ACUs.

This is the key opacity problem. The agent doesn’t pre-quote your task before it starts working. You spin it up, it starts consuming units, and at the end you learn whether it solved the problem or not. Either way, the meter kept running. Replit’s “Effort” credit system has the same problem—you’re flying blind on cost until the invoice lands.

Why does this definition matter? Because it means ACU burn is not proportional to task complexity. Two devs asking Devin the exact same question in slightly different ways will get wildly different invoices. Well-scoped tasks (bug fixes with a clear log line, test writing for a known module) cost less. Vague architectural questions (“redesign our auth flow”) cost a lot more because the agent burns ACUs exploring the codebase and asking clarification questions internally.

The Plan Math: Free / Pro / Max / Teams

Cognition offers four tiers:

Free: Limited quota, decent for trial runs. You’ll hit the limit inside a week if you’re serious.

Pro ($20/month): ~9 ACUs included, $2.25 per overage ACU. The trap tier. Looks affordable until the bill lands.

Max ($200/month): Larger quota, better per-ACU effective rate. More sustainable for solo devs doing regular work. Most active users should start here instead of Pro, then downgrade if they don’t use it.

Teams ($80/month): Centralized billing, multiple team members, admin dashboard. The rates are the same as Pro, but you’re pooling quota and getting organizational tooling.

Do the math before you pick. If you’re a solo dev and you’re planning to use Devin more than once a week, Max is probably cheaper than Pro overage in the long run. A team of three devs should absolutely pick Teams and track quota per person so you don’t get surprised.

Real Monthly Bills: Three Usage Profiles

Let’s math this out:

Light user (script-y tasks, 20–30 ACUs/month):

• 9 ACUs included in Pro
• ~20 additional ACUs used
• $20 + (20 × $2.25) = $65/month

Active solo dev (100 ACUs/month):

• 9 ACUs included in Pro
• ~91 additional ACUs used
• $20 + (91 × $2.25) = $225/month

Or just buy Max ($200/month) and stop worrying. Max includes a bigger quota, so you’re covered and the per-unit cost is lower.

Team on a sprint (300 ACUs/month):

• Teams plan at $80/month (same rate as Pro, but pooled)
• ~291 additional ACUs used
• $80 + (291 × $2.25) = $736/month

None of these is “$20.”

The Gotcha: What Actually Triggers ACU Consumption

The hidden assumption baked into the “$20” headline is that users will write bulletproof prompts every time. In practice, nobody does. TechCrunch’s coverage noted that Devin 2.0 completes tasks more efficiently than 1.0, but efficiency is measured against well-scoped tasks. The cost spiral starts when that assumption breaks.

Failed runs burn ACUs. Re-prompts on vague specs burn ACUs. Asking the agent to navigate a large codebase burns ACUs because it has to spend inference tokens understanding your repo structure. Clarification loops burn ACUs. A task you thought would take one run takes three because the agent hit a dependency it didn’t anticipate.

Everyone learns by firing off ambitious tasks, seeing them fail halfway through, and re-running with more context. That’s the real cost curve.

Devin vs. the Other Agent Billing Models

Devin isn’t alone in this. Every coding agent has punted the cost-transparency problem downstream.

Replit’s “Effort” credit system uses a similar opacity—you don’t know how many credits a task will burn until it’s done. OpenAI’s Codex-era billing was the same trap. Manus is building an agent framework and hasn’t published detailed billing yet, but the pattern is consistent: units consumed, no pre-quote, learn-as-you-go pricing.

The industry has collectively decided that cost visibility is a buyer problem, not a seller problem. None of the agent startups pre-estimate task cost before running. It’s a competitive disadvantage to do so (the estimate might scare you away), so nobody does it.

When Does the $20 Plan Actually Make Sense?

The $20 plan is real for one specific persona: light solo devs doing well-scoped, bounded work. You’re writing a utility script, fixing a known bug with a clear log, writing tests for an existing module. You run Devin maybe 2–3 times a month. You’re not iterating hard. The included 9 ACUs carry you, and you pay $30–40/month in overages.

Above that usage floor, Max at $200/month is a better hedge than stacking overages. You get more quota, lower per-unit cost, and you stop stress-testing your invoice every time you fire the agent.

Teams at $80/month only makes sense if you’re actually organizing around Devin as a core tool and you need the admin dashboard. If you’re just three devs running it ad-hoc, split a Max plan.

Also, check the IDE pricing bundle if you’re considering Devin alongside Windsurf or other cloud IDEs. Cognition is bundling Windsurf access into higher tiers. The total package cost might change your math.

The Real Story

Cognition didn’t lower Devin’s price by 96%. They repackaged the same compute metering behind a lower-looking headline number. The $20 is real. The $200–500 month is also real. It’s the same strategy SaaS startups have used forever: put a low number on the homepage, bury the per-unit cost in the fine print, and let the market figure out what actual usage costs.

We’re Tool Crier because we cry about pricing like this so you don’t have to. Run the math on your expected ACU burn before you commit to a plan. Start with Pro if you’re testing, but budget for Max if you’re shipping.