Devin

Overview

Devin is Cognition AI's software engineering agent. When it launched in March 2024 it was treated as the first agent that could deliver development tasks on its own, and it defined the Agentic Coding category along the way. It behaves more like an engineer in the cloud: assign work in Slack and it reads code, makes changes, runs tests and opens a pull request inside its own cloud environment, finishing with a report that includes screenshots and logs.

In 2026 the focus moved to what happens after work goes out. Devin Oncall arrived on October 15, investigating, triaging and fixing production incidents after an alert, wired into Datadog, Sentry, Dynatrace, PagerDuty, Jira, Linear and ServiceNow. Around the same time Devin Desktop and the CLI added Gemini 4 Argon and Pylon support so support tickets flow through to diagnosis. On October 19 Devin Fusion scored 68 on the Artificial Analysis Coding Agent Index, the highest on that board, using Opus 5.5 as lead with GPT-6 Luna as second, which Cognition says cuts cost per task 31 percent against Claude Code running Opus 5.5.

On October 25 Cognition open sourced Devin Vulnerability Fixes, three MIT licensed skills covering remote code execution bugs in PHP2Shell, Node.js and GitLab, turning security patching into a routine that can be handed to an agent. The model lineup had already widened with Claude Haiku 5.5 and GPT-6.1 Sol, and Memory lets the same context carry across sessions.

Key Features

• Parallel sessions: One developer can supervise dozens of Devin sessions, each running in its own cloud environment. Tasks stay independent and progress rolls up into a board, which suits bulk maintenance work.
• Devin Fusion mixed routing: Scored 68 on the Artificial Analysis Coding Agent Index on October 19, 2026, the highest on that board. Opus 5.5 leads while GPT-6 Luna handles secondary work, and Cognition puts cost per task 31 percent below Claude Code on Opus 5.5.
• Devin Oncall: An AI SRE agent generally available on October 15, investigating, triaging and fixing production incidents once an alert fires. Integration covers Datadog, Sentry, Dynatrace, PagerDuty, Jira, Linear and ServiceNow, with bug reproduction and a fix PR.
• Open source vulnerability skills: Devin Vulnerability Fixes shipped on October 25 with fix-php-vuln, fix-node-vuln and fix-gitlab-vuln, three MIT licensed Agent Skills covering remote code execution bugs in PHP2Shell, Node.js and GitLab.
• FrontierCode and model lineup: An in-house coding model family where FrontierCode 1.1 handles tasks that general models cover at higher cost, bringing ACU spend down to about 1 dollar per task. Claude Haiku 5.5, GPT-6.1 Sol and Gemini 4 Argon sit alongside for matching to task type.
• Memory and workflow integration: Memory carries project context across sessions so new tasks skip the background explanation. Native Slack, Linear, Jira and Pylon integration brings work in and pushes results back, keeping ticket state and pull requests in step.
• Full cloud development environment: A cloud machine that can actually run the code: install dependencies, start services, run tests and drive a browser. Environments are isolated per session and artifacts live on in the pull request.
• Pull requests and review: Work ends with a pull request carrying a change summary, test results and screenshot logs. Review comments return to the same session for follow-up without restating the context by hand.

Use Cases

• Bulk automation of dependency upgrades, security patches and repetitive bug fixes across dozens of parallel sessions.
• First response to production incidents, where Oncall investigates and triages, then delivers reproduction and a fix PR together.
• Known vulnerability remediation for PHP2Shell, Node.js and GitLab, handed to the agent through the open source skills.
• Test coverage backfills and documentation sync, the dull maintenance work that fits a session queue.
• Small to medium feature work driven by Jira or Linear tickets, with status written back automatically.
• Large refactors that need long context, using Memory so multiple sessions share one project background.

Pros

• Devin Fusion scores 68 on the Coding Agent Index at 31 percent lower cost per task than Claude Code on Opus 5.5.
• Sessions run dozens of tasks in parallel, so maintenance work can be queued.
• Oncall connects incident response to Datadog, Sentry and PagerDuty instead of requiring a separate on-call process.
• Three vulnerability skills ship under MIT, so the implementation can be read before deciding how to use it.
• Memory keeps project background across sessions, saving repeated briefing on long refactors.
• Native Slack, Linear, Jira and Pylon integration keeps assignment and acceptance inside existing workflows.
• ACU billing starts around 1 dollar per entry task, so a small pilot comes before a large rollout.

Pricing

Billing runs on ACUs (Agent Compute Units) per usage, starting around 1 dollar per task since FrontierCode launched, with typical medium tasks landing between 3 and 10 dollars. Subscription tiers: Core for individual developers at about 20 dollars per month with a set ACU allowance, Team priced by seat plus an ACU pool, and Enterprise with custom volume, private deployment, SSO and VPN; the earlier 500 dollar per month team tier has been replaced by elastic ACU billing. On model cost, Claude Haiku 5.5 runs about one eighth of Sonnet 5, GPT-6.1 Sol matches GPT-6 Sol at 2 and 10 dollars per million tokens, and cache reads drop 95 percent. Check the official site for current allowances.

Summary

Devin turned the cloud AI engineer idea into a product teams pay for. In 2026 the emphasis shifted from whether it can do the work to how the work gets accepted: Fusion pushed the Coding Agent Index to 68 while cutting cost per task, Oncall plugged response into alerting systems, and the open sourced vulnerability skills made security patching a reusable routine. For teams running development on Slack, Jira and Linear with plenty of deterministic maintenance work, it is among the most mature options; developers who want speed inside an IDE are better served by Cursor or Claude Code.

Version History

• Devin Vulnerability Fixes released as open source with three MIT skills for PHP2Shell, Node.js and GitLab (2026-10-25): Cognition released the free open source Devin Vulnerability Fixes plugin on October 25. It ships three MIT licensed Agent Skills, fix-php-vuln, fix-node-vuln and fix-gitlab-vuln, each covering the remediation path for one remote code execution bug in PHP2Shell, Node.js and GitLab.
• Devin Fusion tops the Artificial Analysis Coding Agent Index at 68 points (2026-10-19): Cognition announced that Devin Fusion scored 68 on the Artificial Analysis Coding Agent Index, the highest mark on that leaderboard. The model uses a mixed orchestration scheme with Opus 5.5 leading and GPT-6 Luna as second, and Cognition claims it beats Claude Code running Opus 5.5 on both performance and price, with 31% lower cost per task. Devin also introduced an efficiency-based auto-approval mechanism and Devin Coach in the same period.
• Devin Oncall GA; Gemini and Pylon integrations (2026-10-15): Launches Devin Oncall, an AI SRE agent that autonomously investigates, triages and remediates production incidents across Datadog, Sentry, Dynatrace, PagerDuty, Jira, Linear and ServiceNow, reproducing bugs and opening fix PRs; adds Gemini 4 Argon to Devin Desktop/CLI (75% off through Oct 27); integrates Pylon to route support tickets into Devin for automatic triage.
• Devin can now use a personal ChatGPT plan (2026-10-09): Cognition announced that Devin works with any personal ChatGPT plan. Once a user connects a Go, Plus or Pro plan to Devin, all GPT model usage is drawn from that plan quota, with the billing documentation updated accordingly.
• Devin adds Claude Haiku 5.5 support (FrontierCode 1.1 score 58.4%, cost approximately (2026-10-07): Devin adds Claude Haiku 5.5 support (FrontierCode 1.1 score 58.4%, cost about one-eighth of Sonnet 5, joining the sidekick lineup of Devin Fusion); adds GPT-6.1 Sol support (available in Devin Desktop and Devin CLI, priced at $2/$10 same as GPT-6 Sol, with cache reads reduced by 95%); launches Memory
• Devin 3 / FrontierCode (2026): Cognition's self-developed FrontierCode coding model launched, significantly improving long-task success rates, reducing ACU unit price to starting at $1/task, and beginning to achieve economic viability for large-scale enterprise deployment
• Devin 2.x (2025): Sessions multi-task parallelism, native Slack/Linear/Jira integration, upgrading from 'single Devin' to 'Devin team'
• Devin 1.0 (2024): World's first 'AI software engineer' sold to enterprises, defining the Agentic Coding category, with a valuation exceeding $2 billion