Contains affiliate links. We may earn a commission — it never changes the verdict. Details
The Stack / Coding
Devin
03 CreationWorth a real trial on a free account first — Cognition replaced the old $20-plus-ACU Core plan with a Free/Pro/Max ladder on July 7, so 'assign a ticket to an AI' now costs nothing to test before you commit to Pro. It shines on well-specified, bounded tasks (migrations, test coverage, dependency bumps) and still wanders on vague ones. Spec tightly, review everything.
The job in the stack
The delegated-engineering seat: an agent that takes whole tickets — bugfixes, migrations, small features — and returns pull requests. Since June 2026 this also covers the interactive IDE (Devin Desktop, formerly Windsurf) — cloud ticket-work and local agent editing now share one Kanban view under the same brand.
Pricing
| Plan | Price | Notes |
|---|---|---|
| Free | $0 | Light quota, limited model access — former Core users were migrated here |
| Pro | $20/mo | Full model access: SWE-2 plus frontier models from OpenAI, Anthropic, Google and others, and leading open models. Devin Cloud included; allowances refresh daily and weekly; extra usage billed at API rates past quota. SWE-2 usage is free on Pro and above, with no end date shown on the pricing page |
| Max | $200/mo | Same feature set as Pro, significantly higher quota |
| Teams | $80/mo + $40/seat | Usage-based, up to 200 users; each full seat gets its own quota and Devin Desktop, plus admin dashboard and centralized billing. Former $500/mo Team plan migrated here at a lower entry point |
| Enterprise | custom | VPC, SSO, dedicated support |
Pricing verified 2026-10-25. Plans change — the source of truth is Devin's own pricing page.
Models & versions
- Devin (Cognition) Autonomous ticket-to-PR workflow: plans, codes, tests, opens the pull request
- SWE-2 Shipped Sept 10, 2026, replacing SWE-1.7 as the flagship — post-trained from Moonshot AI's Kimi K3 base. Cognition's own benchmarks claim it matches Claude Fable 5.1 on FrontierCode at 64% lower cost, and beats SWE-1.7 by 81% on the same cost measure. Live now in Devin Desktop and Devin CLI; Devin Web and Fusion rollout still in progress. Vendor benchmark, not independently verified — treat the cost claims as directional
- SWE-1.7 (superseded) Cognition's prior coding model, shipped July 2026. Replaced by SWE-2 on Sept 10, 2026
- Devin Fusion Announced June 29, 2026 and in preview — a router, not a model. It runs a frontier main agent alongside a cheap sidekick and hands mechanical work down mid-task. Cognition claims frontier-level quality at 35% lower cost on its own benchmark, later reporting up to 60%. It runs on the plan you already have
The honest read
What works
- + Whole-ticket delegation: it returns PRs, not suggestions
- + Excellent at grind work — migrations, coverage, upgrades — that stalls human queues
- + Free tier now lets you evaluate before paying anything
What breaks
- − Vague tickets produce expensive wandering — specification quality is the leverage
- − Every PR still needs human review; autonomy isn't correctness
- − No more pay-as-you-go-only option — ongoing use past Free now requires a paid plan
- − Trust boundary to watch: SpaceX approached Cognition (Devin's maker) about an acquisition in August 2026 — Cognition didn't engage, but talks continue about Cognition using SpaceX's compute. Same pattern as the Cursor–SpaceX deal already noted on that page; nothing has changed for Devin users yet, but a compute or ownership tie to SpaceX would be worth watching for the same reasons.
Use it for
- →Engineering teams with backlogs of well-defined maintenance work
- →Framework migrations and dependency upgrades at scale
- →Test-coverage sprints nobody wants to staff
Skip it if
- ×You want interactive pair-programming — Cursor and Claude Code are that
- ×Your tickets are one-line vibes — fix the specs before hiring an agent
Questions operators ask
?Devin or Claude Code?
Claude Code works with you in the terminal — interactive, steerable, immediate. Devin works without you — assign the ticket, get a PR. Interactive depth versus asynchronous throughput; mature teams end up using the pattern, not the brand: supervised agents for hard problems, delegated agents for grind.
?Does Devin Fusion actually cut my bill?
It changes what your quota buys, not what you pay. Fusion routes the mechanical steps of a task to a cheap sidekick model and keeps the frontier model for the parts that need it, so the same monthly quota stretches further. Cognition's own benchmark puts the saving at 35%, with later reporting up to 60% — vendor numbers on a vendor benchmark, so treat them as a direction, not a guarantee. It's in preview and runs on the plan you already have, which makes it worth turning on and measuring against your own last month rather than arguing about.
?What happened to Devin's ACU pricing?
Cognition dropped the $20-plus-$2.25-per-ACU Core plan and the flat $500/mo Team plan in favor of Free, Pro ($20/mo flat), Max ($200/mo), and Teams ($80/mo base plus $40/seat) — a lower, simpler entry point than the old structure.
?Is Devin the same thing as Windsurf?
As of June 2026, yes. Cognition acquired Windsurf in July 2025 and renamed it Devin Desktop — same editor, same team, converged onto Devin's own pricing. If you're comparing the two today, you're evaluating one product, not two.
What changed
- Freshness sweep: re-read devin.ai/pricing. Prices hold (Free, Pro $20, Max $200, Teams $80 + $40/seat, Enterprise custom). Free SWE-2 use on Pro and above is still listed and the Oct 16 end date is gone. Teams now caps at 200 users, and allowances refresh daily and weekly.
- Freshness sweep: Free/Pro $20/Max $200/Teams $80 + $40 per seat (up to 200 users)/Enterprise ladder reconfirmed unchanged on devin.ai/pricing. Fixed the stale Pro model line (it still named SWE-1.7). The page now says SWE-2 is free to use on Pro and above through Oct 16, 2026. That's a limited window, so test it before the meter starts. Cognition reported on Sept 25 that it had passed $1B in annualized revenue run rate. Independent coverage notes SWE-2 trails the frontier models on Terminal-Bench 4, which is consistent with our 'spec tightly, review everything' verdict.
- News scan: Cognition shipped SWE-2 on Sept 10, replacing SWE-1.7 as Devin's flagship coding model (post-trained from Kimi K3). Cognition's own numbers claim parity with Claude Fable 5.1 on FrontierCode at a fraction of the cost — a vendor benchmark, so read it as a direction, not proof. Live now in Devin Desktop/CLI; pricing ladder unaffected, not re-verified this pass.
- Freshness sweep: Free/Pro $20/Max $200/Teams $80+$40 per seat/Enterprise ladder reconfirmed unchanged against a current third-party pricing breakdown (devin.ai/pricing itself 429'd to automated fetch). No change to pricing, models, or the Fusion router's preview status.
- News scan: Bloomberg reported (Aug 19) SpaceX approached Cognition, Devin's maker, about an acquisition; Cognition declined to engage but talks continue about Cognition using SpaceX's compute capacity. Added as a trust-boundary cons line, framed as reported talks, not a done deal — pricing and ladder not re-verified this pass.
- Freshness sweep: the Free / Pro $20 / Max $200 / Teams $80+$40 / Enterprise ladder holds unchanged. Two model-side items we had wrong or missing. SWE-1.7 shipped in July 2026 and replaced SWE-1.6 as the flagship — trained inside the Devin harness for long-running async work, served via Cerebras at about 1,000 tokens/sec, included on Pro. And Devin Fusion, announced June 29 and in preview, is a multi-model router rather than a model: a frontier main agent with a cheap sidekick taking the mechanical steps, which Cognition benchmarks at 35% lower cost (up to 60% in later reporting). Both are Cognition's own published numbers on Cognition's own benchmark — read them as vendor claims, not independent results. Neither changes the price of a seat; both change what a seat's quota buys.
- Freshness sweep: no material change. The Free / Pro $20 / Max $200 / Teams $80+$40 ladder and the SWE 1.6 flagship re-confirmed by search; self-serve quotas are now token-metered (a March 2026 shift already covered by the 'overage billed past quota' note). last_verified advanced.
- Noted while retiring windsurf.json from the stack: Cognition folded Windsurf into the Devin brand as "Devin Desktop" on June 2, 2026 — Cascade became "Devin Local," running alongside cloud Devin jobs in one Kanban view. Pricing unaffected (Free/Pro/Max/Teams ladder confirmed unchanged against devin.ai/pricing); not counted as this run's freshness-sweep verification.
- Released from staging. Cognition retired the Core/ACU plan for a Free/Pro ($20)/Max ($200)/Teams ($80+$40 per seat) ladder — corrected before release; the ACU-billed Core plan no longer exists.
- Added to the stack (delegated-engineering seat).
Give Devin a job
The delegated-engineering seat: an agent that takes whole tickets — bugfixes, migrations, small features — and returns pull requests. Since June 2026 this also covers the interactive IDE (Devin Desktop, formerly Windsurf) — cloud ticket-work and local agent editing now share one Kanban view under the same brand.
Try Devin →Alternatives with different jobs
Cursor
03 CreationThe hands-on coding seat: an editor where AI agents write, refactor, and navigate the codebase with you.
Claude
02 DecisionsThe thinking seat: long documents, hard reasoning, code, and anything where the words have to be right the first time.
GitHub Copilot
03 CreationThe completion seat: always-on code suggestions in your existing editor, plus agent work metered by AI credits.