Devin Retired Its Pricing Unit for Everyone Except the Customers Who Actually Negotiate

Every "recent news" roundup on AI coding agents still leads with the same line: Cognition cut Devin's price from $500 a month to $20 a month plus $2.25 per Agent Compute Unit. It's a great stat. It's also a year stale — that cut happened in April 2025, not 2026, and it's already been superseded by a second, quieter pricing change that's more useful to understand than the one everyone keeps citing.

On April 14, 2026, Cognition retired ACU-denominated billing for every self-serve customer on Devin. Free, Pro, and Max users no longer see "ACUs" anywhere in their bill. Only Enterprise customers — the ones with a signed order form and a negotiating finance team — still get billed in the unit Cognition invented specifically to meter agent compute. That split, not the 2025 price cut, is the actual story: a vendor built a bespoke metering unit for autonomous-agent work, watched it fail with retail users inside of a year, and kept it alive only where a human was going to read the contract line by line anyway. That's the concrete lesson for any platform — bex included — that has to decide what unit to meter once an AI agent, not a person clicking deploy, is the thing generating the bill.

Devin's Three Pricing Eras​

Here's the corrected timeline, pulled from Cognition's own pricing page, its April 2026 blog post, and contemporaneous reporting:

The 2025 change was the one that made headlines: a 96% cut to the entry price, opening Devin to individual developers who'd never clear a $500/month floor. It also introduced ACUs as the customer-facing meter — a normalized unit covering VM time, model inference, and network bandwidth, priced separately from the subscription.

The 2026 change is the one nobody's writing "recent news" posts about yet, and it reverses the customer-facing part of the 2025 design. Self-serve usage beyond a plan's included quota is now shown and billed in plain dollars. Ask Devin, DeepWiki, and Devin Review — previously free — now draw against that same quota. ACUs didn't disappear; they got demoted to a contract term that only Enterprise customers, who negotiate pricing directly with Cognition's sales team, ever see.

Why Cognition Killed Its Own Meter​

Cognition's own blog post gives three reasons for the April 2026 restructuring, and none of them are "ACUs were mispriced." They're about what a metered unit does to a self-serve buyer's head:

A clearer self-serve path. The old Core/Team split left a canyon between "trying Devin" and "adopting it" — there was no smooth ramp from casual use to team use. Free → Pro → Max → Teams replaces two tiers with four, each a smaller step than the last.

A lower barrier for teams. The old Team plan's $500/month entry point was still too high for teams that wanted to use Devin regularly but weren't ready to commit at that level. Teams now starts at $80/month plus $40 per seat — usage-based, not a $500 floor.

Pricing that matches real compute-heavy products. Ask Devin, DeepWiki, and Devin Review all consume meaningful resources that used to be free; extending metering to them required a unit self-serve users could actually reason about.

An independent analysis published a month before this post (The Butler, May 2026) put the underlying psychology in sharper terms: ACU framing made "the cost of active work feel explicit" in a way that read as a meter spinning while you weren't looking. Replacing "0.4 ACUs" with "$0.90" doesn't change what's billed — it changes whether a self-serve user can do the math in their head without opening a conversion table.

The Enterprise Exception Is the Tell​

If ACUs were simply a bad idea, Cognition would have dropped them everywhere. It didn't. Enterprise contracts still bill in ACUs, at whatever rate is set in that customer's order form.

That split is the actual finding, and it generalizes past Devin: a bespoke, abstracted metering unit is a fine thing to put in a negotiated contract, where a finance team reads the definition once, gets it explained by an account rep, and then never thinks about the unit again because procurement already translated it into a budget line. It's a bad thing to put in front of a self-serve signup flow, where the buyer is doing the ACU-to-dollars math themselves, every session, with no account rep to call. Enterprise buyers tolerate — even expect — a named unit as part of the ceremony of a contract. Retail buyers just want to know what a thing costs before they run it.

What Railway Already Got Right​

Devin had to learn this the hard way because it started from a bespoke unit and walked it back. Railway never had that problem, because it never invented one. Its rate card has been plain dollars-per-resource from the start:

A Railway customer never has to ask "what's a Railway unit worth in dollars" — there is no intermediate unit to convert. The bill is denominated in the same currency the pricing page uses, the invoice uses, and the customer's own budget uses. Cognition spent a year converging on that same design for retail users; Railway priced usage that way from day one, presumably because vCPU-seconds and GB-months are units a developer already recognizes from every cloud bill they've ever read, where an "Agent Compute Unit" is not.

The Question Bex Has to Answer Before Its MCP Server Ships​

Here's where this stops being a Devin story and becomes a platform-design one. Every deploy platform, bex included, has priced compute for a world where a human clicks deploy, waits, and either gets a running service or a build log to debug. That's a bounded, infrequent, human-paced source of billable work.

An AI agent operating that same platform over MCP is not bounded or infrequent. An agent debugging a failing build might retry five times in five minutes. An agent managing a fleet of services might poll status, redeploy, and roll back in a loop that a human would never run manually. The thing generating the bill stops being "a person clicked deploy" and becomes "an agent did a thing, and then did it again, and then did it again" — closer to Devin's autonomous-work model than to a traditional PaaS's human-triggered build.

Devin's two-year detour is the concrete answer to what unit that work should be metered in: not a new one. The build minutes, CPU-seconds, and RAM-seconds a platform already meters for a human-triggered deploy are the same units an agent-triggered deploy consumes — an agent retrying a build five times burns five builds' worth of compute-minutes, not some separately-priced "agent action." The mistake to avoid isn't billing agent-driven work; it's inventing a new customer-facing currency to bill it in, the exact thing Cognition spent a year walking back for its own retail customers. Meter agent-triggered work in the same dollars-per-resource units a human-triggered deploy already uses, and an MCP-driven redeploy loop shows up on the invoice exactly like five human-triggered redeploys would — legible without a conversion table, and consistent with Railway's git-push-metering answer to the same question.

The harder part isn't the unit — it's the frequency control. The Butler's analysis of Devin's change makes the same point about Devin Review's new run-frequency settings (manual only, on PR open, or on every commit): flexibility over when metered work triggers is what keeps a bill legible, not just the unit it's denominated in. A platform that lets an agent redeploy in a loop needs the same kind of governance lever — rate limits or retry caps on agent-triggered actions — that Cognition just had to bolt onto Devin Review after the fact. Bex is in the position to design that in before an agent-driven MCP server ships, rather than retrofitting it the way Cognition had to.

The Corrected Story Is the More Useful One​

The "Devin cut its price to $20 plus $2.25/ACU" headline will keep circulating because it's a clean, dramatic number from a year-old launch. The actual 2026 news — that Cognition quietly killed that same metering unit for everyone except the customers with a contract to negotiate — is less dramatic and more instructive. It's a live example of a vendor discovering, in production, that a metered-compute unit belongs in a negotiated line item, not a signup flow, and it's the exact question a deploy platform pricing agent-driven work has to get right before its own metering ships, not after.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, with agent-operable infrastructure exposed over MCP instead of a bespoke SDK or a bespoke billing unit. Star the repo on GitHub or deploy your first app today.

Sources​

• New self-serve plans for Devin — Cognition
• Devin, the viral coding AI agent, gets a new pay-as-you-go plan — TechCrunch
• Devin 2.0 is here: Cognition slashes price of AI software engineer to $20 per month from $500 — VentureBeat
• Plans and Pricing — Devin
• Billing — Devin Docs
• Devin's Pricing Shift: From Compute Meters to Budget Policy — The Butler.