Devin vs Claude Code in 2026: hand it off, or work with it?

Devin and Claude Code both start at $20 a month but are built for different ways of working. Devin, from Cognition, takes a task and runs on its own in a cloud machine, reachable from Slack, Linear, GitHub or its web app, and now includes Devin Desktop, the editor formerly called Windsurf. Claude Code, from Anthropic, is an agent you work with in your terminal, editor, desktop app or browser, and it asks before acting unless you let it run. On the one independent benchmark that tests both, Artificial Analysis's coding-agent index, Claude Code with Sonnet 5.5 scores 68 and Devin's CLI with Fable 5.1 scores 62, at roughly half the cost per task.

Key takeaways

• Devin's plans since April 14, 2026: Free, Pro $20, Max $200, Teams $80 a month plus $40 per full seat, and Enterprise (Windsurf pricing). Only Enterprise is still billed in ACUs.
• Claude Code's plans: Pro $20 ($17 billed annually), Max $100 or $200, Team seats and Enterprise. There is no free plan.
• Models: Devin mixes models from several labs, Anthropic's included, with Cognition's own SWE-2; Claude Code runs Claude models only.
• Benchmarks: Claude Code scores higher on Artificial Analysis's index, 68 against 62; Devin's CLI costs $7.90 a task against $14.19 and finishes faster.
• Teams: Devin's sessions are visible to your organization and teammates can join one, but neither tool asks the team to approve each change.

How do Devin and Claude Code work?

Devin gives each session its own cloud machine with a shell, an editor, a browser and a desktop for testing. You scope a task in Ask mode, hand it over in Agent mode, and it opens a pull request and fixes review comments (Devin docs). Playbooks and knowledge carry instructions between sessions. Cognition now sells three surfaces on one plan:

• Devin Cloud, the autonomous agent, from the web, Slack, Microsoft Teams, Linear, Jira and Git hosts.
• Devin Desktop, which Windsurf became on June 2, 2026, with a board of local and cloud agents (Cognition).
• Devin CLI, a terminal agent launched in April 2026 that can hand a session to the cloud.

Claude Code runs on your machine or in Anthropic's cloud. By default it asks before editing files or running commands, and you choose how much to approve, up to letting it work through a task on its own. It works in the terminal, VS Code, JetBrains, the desktop app and the browser, and on GitHub it answers @claude in issues and pull requests.

How much do Devin and Claude Code cost?

Devin (pricing):

• Free: a light quota and limited models.
• Pro, $20 a month: models from OpenAI, Anthropic, Google and xAI, Devin Cloud, and extra usage at API prices.
• Max, $200 a month: much higher quotas.
• Teams: $80 a month plus $40 for each full seat, for up to 200 users.
• Enterprise: custom, billed in ACUs, Cognition's unit for compute, model use and bandwidth.

On April 14, 2026, Cognition retired its pay-as-you-go Core plan, priced per ACU, and its $500-a-month Team plan; self-serve usage is now billed in dollars (Cognition).

Claude Code comes with Claude plans: Pro at $20 a month or $17 billed annually, Max at $100 or $200, Team at $20 to $25 a standard seat or $100 to $125 a premium seat, and Enterprise at $20 a seat plus usage at API rates. Limits are shared with Claude chat (Claude Code usage limits).

Which is better? The benchmarks

Only Artificial Analysis tests both agents. It compares Claude Code with Sonnet 5.5 against Devin's CLI running Fable 5.1 with Cognition's SWE-2 (Artificial Analysis):

• Overall index: Claude Code 68, Devin 62.
• DeepSWE: 72% against 63%.
• Terminal-Bench 4.0: 66% against 56%.
• SWE-Atlas Q&A: 67% against 66%.
• Cost per task: $14.19 against $7.90.
• Time per task: 1.5 hours against 36 minutes.

This tests Devin's CLI, not the autonomous cloud agent, which no independent benchmark covers. Devin doesn't appear on Terminal-Bench's own leaderboard, where Claude Code with Opus 5.5 leads at 64.8% (the full rankings).

Which models do they use?

Devin mixes models: its Normal and Ultra modes name Anthropic's Fable 5.1 and Opus 5 and OpenAI's GPT-6 Astra, and its own SWE-2, released September 10, 2026, is post-trained from Moonshot's Kimi K3 (Cognition). Claude Code runs Claude models only: Opus 5.5, Sonnet 5.5, Fable and Haiku.

How do they work for a team?

• Devin: sessions are visible to your organization by default, teammates can join a session and answer Devin's questions, and a Slack thread syncs with its session both ways. Devin can also open a dedicated Slack code channel for one task. Review happens on the pull request, not change by change.
• Claude Code: sessions belong to one person; a cloud session can be shared as a read-only snapshot. On Team and Enterprise, Claude Tag shares a session with a Slack channel, running without per-action prompts.

Both companies point to the same pattern: hand the agent a task, review the result. Poly (usepoly.co) takes the other approach for teams that want a say along the way: a browser room where everyone works with one Claude Code or Codex agent, sees each step live, and any member can approve a change before it's made. Free to start. Who approves a coding agent's changes?

Which should you choose?

• You want to hand off tickets from Slack or Linear and review pull requests later: Devin.
• You want to work through problems with the agent in your own repository: Claude Code.
• You want both: Devin documents handing a Claude Code or Codex session over to Devin Cloud (Devin docs).

Cognition, which bought Windsurf in July 2025, raised more than $2 billion at a $48 billion valuation on September 8, 2026 (Cognition). For OpenAI's agent, see Claude Code vs Codex.