Devin
Cognition's autonomous coding agent, now bundled with the desktop IDE that used to be Windsurf, aimed at teams offloading well-specified tickets to background runs.
Overview
Devin runs as a background agent with its own sandbox: a shell, a browser and an editor, wired into Slack, Jira, GitHub and Linear so a task can start life as a ticket and end as a pull request. Cognition also trains the models underneath it. SWE-1.7 shipped 8 July 2026 with vendor-reported scores of 77.8% on SWE-bench Multilingual and 81.5% on Terminal-Bench 2.1. Self-reported, on their own evaluation setup, which is worth remembering before anyone quotes those figures in a business case.
The acquisition tidy-up matters more than the model numbers for most buyers. Cognition bought Windsurf in 2025 and then retired the brand: existing installs updated themselves to Devin Desktop on 2 June 2026, Cascade reached end of life on 1 July and Devin Local replaced it as the default on-machine agent. If your team standardised on Windsurf, you have already been migrated whether you noticed or not. Plans, settings and extensions carried over intact, which is better than these transitions usually go.
Pricing now runs Free, Pro at $20, Max at $200, Teams at an $80 base plus $40 per seat and Enterprise on request. Enterprise still meters in Agent Compute Units, roughly fifteen minutes of agent work each, and the rate is not published anywhere, so you are negotiating against a number you cannot independently check. The realistic use is queue clearing: dependency bumps, test backfill, migration chores and the tickets nobody volunteers for. Do not buy it expecting to replace an engineer, buy it expecting to drain a backlog.
Key Features
- ✓ Background agent with its own sandbox: a shell, a browser and an editor, triggered from Slack, Jira, Linear or GitHub
- ✓ SWE-1.7, Cognition's in-house coding model, released 8 July 2026
- ✓ Devin Fusion, a routing layer that runs a frontier main agent alongside a cheap sidekick model and switches between them mid-task
- ✓ Devin Desktop, the IDE previously sold as Windsurf, with Devin Local as the bundled on-machine agent
- ✓ Devin Cloud agent access starting on the $20 Pro plan
- ✓ Enterprise controls including SSO, audit logs, SOC 2 attestation and US or EU data residency
Where it holds
- • Entry price fell from $500 a month to $20, which finally makes a real pilot cheap enough to run without a budget conversation
- • Ticket-to-pull-request flow through Slack and Jira matches how teams already queue work
- • Cognition trains its own models and routing, so cost per task has been dropping faster than frontier API prices
Where it breaks
- • Enterprise billing is in Agent Compute Units with no published rate, which makes cost comparison against a $19 Copilot seat guesswork
- • Success rate falls off sharply on ambiguous or architectural tickets, and it rarely signals that it is unsure before producing the wrong thing
- • Headline benchmark numbers are self-reported on FrontierCode, a benchmark Cognition built and runs itself
- • Teams at $80 base plus $40 per seat is steep for a small squad sitting next to much cheaper assistant seats
My Take
The failure mode is specification, not capability. A ticket with acceptance criteria and a test to satisfy will often come back as a working branch; something architectural or loosely worded produces a confident wrong answer with no uncertainty flagged. What changed in 2026 is price and packaging: entry fell from $500 a month to $20, Windsurf became Devin Desktop in an over-the-air rebrand on 2 June 2026 and Devin Fusion now pairs a cheap sidekick model with the frontier one to cut cost by about 60%. That 60% is measured on FrontierCode, a benchmark Cognition built and runs, so discount accordingly.
Quick Info
- Pricing:
- freemium
- Openness:
- Proprietary
- Starting at:
- Free tier with limited quota and model choice. Pro $20/month, Max $200/month, Teams $80/month base plus $40 per seat, Enterprise on request. Entry pricing fell from $500/month during 2026. Earlier subscribers are grandfathered at $15 Pro and $30 per seat. Self-serve plans run on token-based quotas with overage at model API rates; Enterprise still bills in Agent Compute Units, roughly fifteen minutes of agent work each, and Cognition does not publish the per-ACU rate.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Devin holds and where it breaks. The fastest check is your own workload.
Visit website ↗Devin compared head to head
Alternatives to Devin
Cursor
freemiumA VS Code fork built around agentic editing, with its own Composer models, a pull request review bot and SOC 2 controls for teams that want AI in the editor.
GitHub Copilot
freemiumMicrosoft's AI coding assistant across VS Code, JetBrains and the GitHub web, with background agents that open pull requests and org-level policy controls.
Trae
freemiumByteDance's AI IDE and standalone SOLO agent that scaffolds projects, runs terminal commands and tests in a browser, at $10 a month for Pro against Cursor's $20.
Qwen3-Coder
open-sourceAlibaba's open-weights coding model family plus the Apache 2.0 Qwen Code terminal agent, separate artifacts from the hosted Qwen Chat product. Qwen3-Coder-Next runs 80B total, 3B active.