The economics of the price pivot
Cognition slashed the entry price of Devin from $500 per month to $20 in April 2025. This 96% reduction did not trigger a revenue collapse, as the company reported a 13x revenue increase to a $492 million annual run rate a year after the change. While the headline price fell, revenue flows through Agent Compute Units (ACUs). A moderately complex code refactoring task uses 5 to 20 ACUs at $2.00 to $2.25 per unit. Companies using Devin on production codebases often spend $1,000 to $5,000 in ACU costs per month on top of seat fees. While the $500 monthly price tag initially targeted engineering leaders who needed to evaluate if the autonomous category was real, the pivot to a $20 entry point allowed Cognition to capture the massive volume of tasks sitting in enterprise backlogs.
One ACU corresponds to a mid-complexity task where the agent repeatedly reads code and plans. During these sessions, the agent burns between 100,000 and 1,000,000 tokens per task. At the $500 level, one ACU costs $2. The underlying Claude Sonnet token costs between $0.50 and $2.00, which keeps the gross margin for Cognition tight.
Task delegation vs productivity multipliers
The market distinguishes between IDE-integrated tools and autonomous agents. Tools like GitHub Copilot and Cursor are productivity multipliers that work alongside developers. These tools scale with developer headcount. In contrast, Devin is a task delegation product. Revenue for Devin scales with the total backlog of work an organization wants to delegate. You, who already track these LLM benchmarks, know that a demo is a curated sample.
The two categories of AI coding tools serve different economic purposes. IDE tools increase the speed of individual developers by keeping humans in the loop for every keystroke. Enterprise customers pay per seat to gain more output per salary dollar. Devin targets the work that stays in the backlog, such as dependency upgrades or test suite maintenance. Mercedes-Benz used the agent to compress an eight-month legacy system modernization project into eight days. Itau resolved 70% of its security vulnerabilities automatically using Devin.
| Product | Mode | Pricing | Autonomy Level |
|---|---|---|---|
| Devin | Remote agent | $20/month + ACUs | High |
| Cursor | In-IDE agent | $20 to $200/month | Medium |
| GitHub Copilot Workspace | In-GitHub agent | $39+/month | Medium |
| Claude Code | CLI agent | Included in Pro/Max | Medium |
| Replit Agent | Browser agent | $25+/month | Medium-high |
The customer list includes Goldman Sachs, Citi, Dell, Santander, Palantir, NASA, and units of the US Army and Navy. These users signal the shift from experimental testing to production integration. The company’s revenue trajectory drove its recent $48 billion valuation, up from a $26 billion valuation in May 2026. This follows a $1 billion Series D.
Deployment realities and the demo gap
The gap between an agent demo and agent production is structural. An agent remains confident whether it is right or wrong. A human junior engineer who is lost mostly looks lost. An agent that is lost looks exactly like an agent that is succeeding because it maintains fluent narration and tidy diffs until someone checks the work. This failure mode led Devin to spend days in technical dead ends during independent testing. The Register reported in January 2025 that an independent evaluation of twenty real tasks yielded only three satisfactory results.
Reliability remains the primary hurdle for widespread enterprise adoption. In March 2024, Devin claimed a 13.86% success rate on the SWE-bench dataset, but this figure only represented a 25% subset of the benchmark. Critics, including the YouTube channel Internet of Bugs, identified instances where the launch video overstated what the agent achieved on Upwork tasks. The analysis found that the viral demo involved a task that asked for help running an existing computer-vision model rather than writing code from scratch. The video showed Devin fixing its own bugs, but some of the files it fixed did not exist in the original repository. However, Cognition’s internal adoption provides a different data point. In December 2025, Devin wrote 13% of the company’s code, but that number reached 89% by mid-2026. Does the extreme concentration of engineering work in autonomous agents eventually starve the human workforce of training opportunities?
