Devin

  • Devin — Cloud-Sandboxed Autonomous AI Software Engineer That Takes a Ticket to a Merged PR (Cognition, ex-Windsurf Now Devin Desktop, Slack/GitHub/Linear-Driven, Not an In-Editor Copilot)

     

    1. One-line positioning

    Devin (Cognition, SF, Aug 2023; Scott Wu + Steven Hao + Walden Yan; $1B+ raised @ $26B post-money May 2026; acquired Windsurf Jun 2026 → rebranded Devin Desktop) = the cloud-sandboxed autonomous AI software engineer that takes a whole ticket to a merged PR — plan → clone repo → write code → run tests → self-debug CI → open PR → respond to review comments, driven from Slack/GitHub/Linear/Jira, not inside your editor. Not Cursor (AI IDE, interactive), not Claude Code (terminal agent, steerable), not GitHub Copilot (autocomplete+agent plugin), not Factory AI (same autonomous lane, $100 Plus middle tier).  fits ai-productivity-data; pipeline_stage none (spans code-generation + refactor + test-gen + migration + pr-review-respond). Table-first, Claude Code skeleton. The 2026 trap: April-2026 reprice killed the old $500 Team + ACU-meter​ — new self-serve is Free/Pro $20/Max $200/Teams $80+multi-seat, but overage bills in dollars at model API rates, not ACUs; legacy ACU ($2.25/≈15min) still lingers in Enterprise contracts and stale aggregator pages. That’s the line to encode.

    2. Core features

    Layer Capability Notes
    Runtime Own VM sandbox per task — private shell, browser, editor, git, package install Never runs on your machine; async minutes–hours
    Autonomy Ticket→PR: plans, explores repo, edits across files, runs test suite, fixes failures, rebases, opens PR, replies to PR comments Dynamic re-planning on roadblock (since 3.0)
    Models Default Cognition SWE-1.7 (free on Pro, Cerebras ~1k tok/s) + frontier pick OpenAI/Claude/Gemini + Devin Fusion router (cheap sidekick + frontier main) SWE-1.7 behind Opus 4.8/GPT-5.5 on FrontierCode
    Parallel Spawn 5–8 concurrent agents (migrations, per-service fan-out) >5 concurrent → coordination overhead
    Integrations Slack @Devin, GitHub/GitLab PRs, Linear, Jira, MCP server, REST API, Devin Desktop (ex-Windsurf) No JetBrains/Vim core
    Enterprise VPC-isolated deploy, SAML/OIDC SSO, SCIM, audit logs, Devin Security Swarm (separate SKU), SOC 2 Type II + ISO 27001 FedRAMP High in progress
    Surface 🌐 Web console · 💬 Slack/Teams · 🖥️ Devin Desktop (mac/win/linux) · 🔌 IDE ext (VS Code) · REST API + MCP No mobile editing
    Data Paid tiers can opt out of training; Free tier defaults to allowed-training (same pattern as DeepL free) —

    3. Pricing (2026-08, post 2026-04-14 reprice, USD/mo)

    Tier Price Seats / quota Unlocks
    Free $0 light agent quota, limited models (SWE-1.7 only), unlimited inline/Tab in Devin Desktop Try ticket→PR, no frontier models, no DeepWiki
    Pro no annual discount published) higher quota, full frontier pick, Devin Cloud agents, SWE-1.7 free, overage at API $ Solo devs / consultants
    Max $200 significantly higher quota than Pro, same features Heavy solo agent users
    Teams $80 base + $40/full dev seat/mo (flex seats free) shared billing, admin dashboard, priority support, centralized KB 2+ dev teams
    Enterprise Custom VPC, SSO, ACU-based or commit-based, Security Swarm, dedicated account team Banks/government (Goldman, Santander, NASA)

    Legacy (retired 2026-04-14, still on stale pages): Core $20 + $2.25/ACU; Team $500 + 250 ACU @ $2.00. New overage = dollar-based at model API pricing, not ACU.​

    pricing_type = freemium; starting_price = $0; free_tier = "Free: light agent quota, SWE-1.7 only, unlimited inline/Tab in Devin Desktop, no frontier models, no DeepWiki, training-opt-out NOT available"; pricing_note = "apr-2026-reprice-killed-acu-self-serve; overage-now-dollar-at-api-rates-not-2.25-acu; legacy-core-20-plus-2.25-acu-and-team-500-stale-pages; teams-80-base-plus-40-per-full-seat-flex-free; max-200-no-middle-100-like-factory; free-tier-trains-models-paid-opt-out; no-annual-discount-published; security-swarm-separate-sku"

    4. Access Type (v2)

    access_type: web-app, desktop-app, api

    access_display: 🌐 Web console · 💬 Slack/Teams · 🖥️ Devin Desktop (ex-Windsurf, mac/win/linux) · 🔌 VS Code ext · REST API + MCP

    5. Reviews (split — G2/Trustpilot divide)

    Platform Score Volume Signal
    G2 4.6/5​ ~5 (thin) Ticket→PR autonomy, migration wins, Slack UX
    TheAIAgentIndex 4.6/5​ — Autonomy/integrations high, pricing clarity 4/5
    Trustpilot ~3.0/5​ very small Billing surprise on ACU overage (legacy), refund friction
    Aggregate ~4.5/5​ (G2+index) — Trustpilot gap = legacy ACU overage shock, not product quality; same shape as Cursor/Ahrefs

    Recurring praise: Nubank 8–12x ETL migration speed, Mercedes 8-month→8-day legacy cut, Itaú 70% auto security-fix, parallel agents for dep-bumps/test-coverage grunt work, Slack @Devin ergonomics.

    Recurring gripes: ~15% completion on hard ambiguous SWE-bench-style tasks, vague tickets → expensive wandering, silent mid-task stalls, async minutes–hours kills pair-programming loop, ACU-era bills blew past $500, free-tier trains models, needs precise specs + human review on every PR.

    6. Best for / Not for

    Best for Not for
    Eng leads with backlog of well-scoped grunt: framework migrations, dep upgrades, test-coverage sprints, boilerplate CRUD Interactive pair-programming / tight edit-review loop (Cursor/Claude Code)
    Enterprise legacy modernization (COBOL→Go, monolith→services) with CI + good specs Open-ended architectural design / senior judgment calls
    Teams on Slack+Linear+GitHub wanting async ticket-takers overnight Vague “build me an app” prompts (fails silently, burns quota)
    Agencies doing repeatable client migrations at volume Budget-predictable small teams (dollar overage unpredictable vs $20 flat Cursor)
    Orgs needing VPC + SSO + audit (banks, gov) JetBrains/Vim-only devs (no native port)
    Non-devs in Slack who can spec a bug clearly Anyone expecting senior-engineer correctness without reviewing PRs

    7. Competitors (table)

    Tool Lane Starts vs Devin
    Cursor AI-native IDE + background agents $20/mo Interactive, cheaper flat seat, no ticket→PR autonomy, no Slack async
    Claude Code Terminal agent (CLI) $20 Pro / $200 Max Steerable, 1M ctx, best SWE-bench (~87%), no cloud sandbox PR machine
    GitHub Copilot IDE plugin + agent mode $10/mo Cheapest entry, GitHub-native, shallower autonomy, no parallel ticket spawn
    Factory AI Autonomous SWE agent $20 Pro / $100 Plus / $200 Max Same lane, has $100 middle tier Devin lacks, Droid + Codegen split
    OpenAI Codex CLI Terminal coding agent API-priced No UI/Slack/PR machine, pairs under Devin sometimes
    Windsurf AI IDE (now Devin Desktop under same parent) $15/mo Folded into Cognition Jun 2026; Cascade agent < Devin cloud autonomy
    Augment Code Repo-aware coding assistant custom Stronger RAG context, not full autonomous PR owner

     

Do Not Sell or Share My Personal Information Cookie Settings