Cognition holds a $48 billion valuation and reported $900 million in annualized run-rate revenue. The company signed a multi-year Strategic Collaboration Agreement with AWS on September 15, 2026, to address legacy system modernization. Devin follows a three-step loop where it analyzes existing system behavior, writes replacement code, and tests the new implementation against known outputs to confirm parity. Mercedes-Benz, Echo Global Logistics, and ActiveCampaign use Devin through the AWS Marketplace. Nubank used Devin to migrate 6 million lines of code across 100,000 data class implementations, which produced 12x efficiency gains in engineering hours and 20x cost savings. This migration reduced the time for a single sub-task from 40 minutes to 10 minutes. The February 2026 update allows engineers to run multiple Devin instances simultaneously to handle parallel tasks. This update also improved context retention for tasks that involve many files. Devin acts as an autonomous agent that replaces a task on a sprint board. It integrates with Slack, GitHub, GitLab, Linear, and Jira. Unlike coding assistants such as Cursor or Copilot, Devin handles the entire task lifecycle from planning to pull request. Devin Desktop works on Mac, Windows, Linux, and JetBrains platforms. The company provides an enterprise tier starting at $500 per month for companies that require dedicated commercial terms. However, Devin fails when tasks require novel architecture decisions or security-sensitive code rewrites.
Poolside targets the model layer with Laguna
Poolside maintains a $3 billion valuation and discusses a $500 million funding round led by Bain Capital Ventures. The company focuses on the model layer for coding tools using the Laguna model family. Laguna-xs-2.1 handles page reading and tool calls, while Laguna-s-2.1 fails to return valid JSON even when it receives a schema parameter. Laguna-s-2.1 also proves too slow for page-sized prompts, and it often hits request timeouts during long sessions. Explorbot testing showed that Laguna-xs-2.1 can successfully plan a test suite and execute browser actions. Users can access Laguna-xs via OpenRouter at rates of $0.06 for input and $0.12 for output per million tokens. This pricing is lower than the costs associated with Groq for similar tasks. The Laguna-xs model uses the model slot to read pages by processing cleaned HTML and an ARIA tree. The company also publishes its model weights on Hugging Face. Poolside is currently offering its endpoint for free for a limited time. You should consider Poolside if your workflow requires an open-weight model to power a custom agent.
Choosing between an agent and a model
Devin provides a complete IDE through Devin Desktop, which rebranded the Windsurf environment to include an Agent Command Center. This interface manages local and cloud agents using the open Agent Client Protocol. If you manage a team of engineers, you will find that the Devin Teams plan requires an $80 monthly minimum to support multiple users and shared credit pools. Devin’s pricing for self-serve users covers Free, Pro, Max, and Teams tiers.
| Category | Devin (Self-Serve) | Poolside (Laguna-xs) |
|---|---|---|
| Entry Price | $20/month | $0.06/1M input tokens |
| Deployment | Cloud or VPC | API or VPC |
| Primary Use | Autonomous tasks | Model API |
Devin is the superior choice for enterprise teams that need to automate complex, well-scoped engineering tasks like bug fixes or data migrations. The tool handles the mechanical work of writing and testing code within a sandboxed environment. Devin’s capability to automate the migration of 100,000 data class implementations while reducing the time for a single sub-task from 40 minutes to 10 minutes provides a massive advantage for large-scale enterprise refactoring projects that once took years. Poolside is the better choice for organizations building custom agentic frameworks that require low-cost, open-weight models for specific tasks like reading web pages. Devin provides an enterprise tier starting at $500 per month for companies that require dedicated commercial terms. Does the lack of structured output in Laguna-s-2.1 prevent it from ever serving as a planning agent?
