Devin: The Autonomous AI Redefining Software Development

Devin, the AI Agent Disrupting Software Development

Launched in June 2024 by the American start-up Cognition, Devin is an artificial intelligence agent that aims to replace human developers. With a valuation of $10 billion, Cognition offers a tool capable of autonomously and asynchronously performing complex software engineering tasks. Unlike other solutions like Claude Code or Codex, Devin positions itself as a true virtual software engineer, already attracting giants like Microsoft, Goldman Sachs, and even the U.S. military.

Devin is designed to carry out high-value software development tasks, and its experimental approach has already caught the attention of several major companies. The agent stands out for its ability to operate asynchronously, minimizing the necessary human interactions. When a task is assigned to it, the agent provisions a virtual machine in the cloud, clones the source code repository, installs dependencies, and gets to work completely independently. It can read technical documentation, plan architectures, write code, execute tests, and submit pull requests, with human intervention only at the beginning and end of the process.

Autonomy at the Heart of Devin

Cognition has developed its own models, the SWE family, to optimize Devin's performance. The SWE-1.6 model, launched in April, is trained via reinforcement learning in real-world environments and operates at a speed of 950 tokens per second, which is 13 times faster than Anthropic's Sonnet 4.5. SWE-grep, a specialized model, allows for rapid contextual searches in large code repositories.

In practice, Devin is capable of handling a variety of use cases, from large-scale code migration to the development of new features, as well as automatic pull request reviews and bug fixes from tickets. This versatility makes it a valuable tool for companies looking to optimize their development processes.

A Premium Business Model

Cognition has opted for a pricing strategy focused on profitability. Devin is billed in ACU (Agent Compute Unit) units, which aggregate virtual machine time, model inference, and network bandwidth. One ACU corresponds to approximately 15 minutes of active work. Three plans are offered:

• Core: pay-as-you-go, minimum $20 per month, at $2.25 per ACU.
• Team: $500 per month for 250 ACUs, with access to the API and unlimited parallel sessions.
• Enterprise: custom pricing for large accounts with specific needs.

The hourly cost of Devin ranges between $8 and $9, a rate comparable to that of a human developer or an offshore contractor. The Team plan, with its 250 monthly ACUs, offers the equivalent of about 62 hours of guaranteed autonomous work, according to Cognition.

Real-World Testing

To evaluate Devin, we subjected it to a project for developing a smart health application based on generative AI. Using the GPT-5.4 model, the agent generated a web interface that allows users to ask health questions and send images for analysis. In less than 10 minutes, the project was delivered, and after a few minor adjustments, it was ready for production in under 20 minutes. Devin used 3.45 ACUs for this session, costing €7.76 with the Core plan.

We used Devin's web interface connected to our GitHub account to generate the project. Devin offers two modes: a classic agent mode and a Fast mode. We chose the former to limit token consumption. The prompt given to the agent was as follows:

HealthChat – AI-Generated Health Web Interface

Generate a health web interface using generative AI. The user can:

• Ask an AI health questions via chat
• Send images (X-rays, body photos, etc.) for visual analysis
• Simple, modern, user-friendly interface
• Mobile-first (usable on smartphones)
• Messaging style: user/AI bubbles, input field + button for images
• Permanent disclaimer at the bottom: "⚠️ This tool does not replace medical advice. In case of emergency, call 15."

AI Model Configuration

• Default: OpenAI API key, model gpt-5.4 (with vision support for images)
• Alternative: access to an open-source model via Ollama (server URL + model name configurable)
• Parameters are accessible via a ⚙️ button

Onboarding (First Connection)

Upon first use, ask the user to provide their quick variables:

• Medical history (ATCD/MHT)

This information is stored in localStorage and can be modified at any time.

Prompt System

The user profile (first name, age, weight, history) is automatically injected into the prompt system at each exchange. The prompt system must require the AI to:

1. Ask clarification questions to the user before responding, limited to 5 questions maximum, to gather as much context as possible
2. Never make a diagnosis, only inform and guide
3. If an image is attached, describe what is observed without interpreting it as a diagnosis

Technical Constraints

• Local storage only (localStorage)
• Deployable statically (Vercel, Netlify, GitHub Pages)
• Images converted to base64 before API submission

During generation (in the background), it is possible to follow Devin's actions live from its VM in GUI mode. The experience is quite striking.

Once the code is fully generated, Devin tests both the mobile and desktop versions of our application in its interface. Again, it is quite surprising: the agent seems truly excellent in computer use.

In less than 10 minutes, the entire project is delivered. Aside from a few minor tweaks via a prompt (adding a README, changing a model, adding a license file...), the final project is operational and production-ready in under 20 minutes total. A true record. Devin's real strength lies in not overwhelming the user with various questions and developing quickly and quietly in the background. Devin used 3.45 ACUs for this session, with a real cost of €7.76 under the Core plan (excluding the monthly subscription of $20). The cost is indeed high, but the final application works perfectly. The installation from the repository goes smoothly, as does the usage in real-world conditions.

An Agent That Delivers on Its Promises

At the end of this test, Devin stands out as a credible, and even formidable, alternative to the code agents we have evaluated in recent months. While alternatives to Claude Code, Codex, or Gemini each have their merits, it is clear that few have delivered results as convincing as the market leaders in real-world conditions.

The true differentiator for Devin can be summed up in one word: autonomy. Devin takes the logic further with its dedicated virtual machine. The experience is smoother, less verbose, and the results reflect that. The delivered code is nearly production-ready, particularly because the agent seamlessly combines development, testing, and visual verification—a reflex that only the most seasoned users of Claude Code think to demand from their agent.

All of this comes at a price, and it is far from negligible. With an effective hourly cost between $8 and $9, Devin is significantly more expensive than its direct competitors. However, there is an indirect advantage to this transparent pricing: Devin's price closely reflects the actual cost of generative AI when it is intensively mobilized for complex tasks. Companies adopting this tool today are acclimating to a billing level that will, sooner or later, be that of the entire market. When the time comes to pay AI its true value, the transition will be much less abrupt.