● aa index
—
Pair not measured
Starting at
from $0
Free - a light quota in Devin Desktop; cloud Devin from Pro at $20/mo
Surfaces
6+
cloud · desktop · cli · slack-teams · tickets
Latest release
Sep 21
From the card's track record
Quick start · Web (cloud)
$ → app.devin.ai▌benchmarks and competitors - analysis in 3.0 ↓
1.0
Should you get it
Use it if you need
- ✓batches of delegable tasks: bug fixes, migrations, small features - in parallel in the cloud
- ✓teams with a review process: Devin’s output - a PR approved by a human
- ✓delegation from Slack/Teams, Jira, and Linear - the developer’s machine stays free
- ✓enterprise with VPC deployment, roles, and SSO requirements
Choose it for delegating batches of tasks to the cloud with review at the end; not as your only daily agent
Pass by if needed
- ✗daily hands-on work in your own repo - the terminal-agent category (Claude Code, Codex)
- ✗predictable budget: message pricing varies, with no official formula
- ✗
- ✗open-source bet: the core is closed, with no product repository
Free $0 - a light quota in Devin Desktop, no cloud. Cloud Devin starts at Pro $20/mo, Max $200/mo, Teams from $80/mo.
2.0
Installation by surface
Devin Cloud (app.devin.ai)→
The main surface: a task goes to a cloud machine with a browser, terminal and editor, and comes back as a PR. Machines run on Linux, Windows and macOS. Windows costs about 9% more than Linux, macOS still costs the same as Linux. Pro and Max allow up to 10 parallel sessions, Teams has no limit.
→Connect a GitHub repository and describe the task with acceptance criteria
>_Devin Desktop (formerly Windsurf)→
A VS Code-based IDE that Windsurf turned into on Jun 2, 2026. Inside it runs the local agent Devin Local, and the Agent Command Center panel also drives cloud sessions. Since Sep 11, 2026 Desktop has a Fusion mode.
>_CLI→
A terminal agent on the same engine as Devin Local. You pick the model, there is a Fusion mode, and the /handoff command sends a long task to the cloud. Over the ACP protocol the same agent works inside JetBrains, Zed and Xcode.
$ curl -fsSL https://cli.devin.ai/install.sh | bash$ devin>_Slack and Microsoft Teams→
You tag @Devin in a thread, the task goes to a cloud session, and the answer comes back to the thread. The main delegation path for teams, available from Pro.
→Connect the integration in Settings: Integrations, then Slack / Teams
→Tag @Devin in a thread with a task description
>_Linear, Jira and automations→
You assign a Linear or Jira ticket to Devin and it turns into a PR. Automations start sessions on their own from Slack, GitHub, Linear and PagerDuty events, on a schedule or a webhook.
→Connect Linear or Jira in Devin settings
→Assign the ticket to Devin or mention it in a comment
>_API→
Starting sessions programmatically: a queue of tasks from your own systems (tickets, crons, CI) without the web interface. In the pricing table the API is open from Pro.
What you need before starting
You need a devin.ai account and GitHub connected.
3.0
Rankings and benchmarks
No Devin+model pair in the site ranking · card reference points
AA Index · Devin Fusion CLI + Fable 5.1 and SWE-2
61.7 · 09.2026 · snapshot of Sep 25: third place in the overall AA list, level with Codex + GPT-6 Astra (61.65). A task costs $7.90, while the leader, Claude Code + Opus 5.5, scored 66 at $13.04
AA Index · Devin Fusion CLI + GPT-6 Astra and SWE-2
58.9 · 09.2026 · cheaper: $4.54 per task and 25 minutes against 36 for the pair with Fable
SWE-bench (unassisted)
13.86% · 03.2024 · the only first-party measurement of cloud Devin
Who is measured and how it is calculated - AI agent ranking
● agent ≠ model
An agent's score is counted for an "agent + model" pair. Artificial Analysis measured Devin Fusion CLI, the terminal version of the same agent, while cloud Devin was not measured separately. Devin is not in our AI agents rating yet.
4.0
What it can do: autonomy, techniques, character
Autonomy axis
Devin's autonomy axis goes further than anyone else's: not an agent in your terminal, but a fleet of sessions in someone else's cloud. The price is acceptance: everything returned is checked by a human.
Devin· covers the entire axis
01Task setup
task from the web, Slack/Teams, Jira, or API - the developer's machine stays free
02Autonomous session
its own VM with browser, terminal, and editor; Devin plans and executes on its own
03Fleet
parallel sessions: a batch of tasks goes to the cloud at once
04Acceptance
output - PR; results are not accepted without human review
Asks at every stepWorks while your laptop is closed
Project character · releases up, incidents down
2026-09-25
policy
Cognition's annualized revenue reached $1B - In May it was $492M.
2026-09-21
release
SWE-2 in cloud Devin - Cognition's own model showed up in the agent's model picker as a research preview.
2026-09-11
release
Fusion in Devin Desktop and the CLI - The same day Artificial Analysis added Devin Fusion CLI to the index: the Fable 5.1 + SWE-2 pair scored 61.7.
2026-09-10
release
Its own SWE-2 model - On Cognition's own FrontierCode 1.1 benchmark it gets 50.0% against 50.9% for Fable 5.1, at a 64% lower price.
2026-09-08
policy
A round of more than $2B at a $48B valuation
2026-06-29
release
Devin Fusion in cloud Devin - The lead model and the cheap helper work in parallel. Per vendor data from Aug 7, 2026, on FrontierCode this is up to 60% cheaper at the same level.
5.0
Pricing and limits
Plans · $/mo
free
Free
Devin Desktop, no cloud
$20
Pro
cloud Devin, up to 10 sessions at once
$200
Max
weekly quota without a daily cap
$80
Teams
the team minimum, $40 seats are included in it
Plans
Free $0 · Pro $20/mo · Max $200/mo · Teams from $80/mo
Free gives a light quota in Devin Desktop without cloud Devin. Pro at $20 opens cloud Devin, Claude, GPT, Gemini and Grok, up to 10 parallel sessions. Max at $200 gives noticeably more quota, weekly only. Teams costs at least $80/mo, a full seat is $40 and counts toward that minimum. Enterprise is billed in ACU under a contract. Beyond the quota you buy credits at API prices, and they do not expire.
Devin pricing
2026-09-26
Free access
Yes, but only Devin Desktop with a light quota
Per the pricing page, cloud Devin starts with Pro at $20/mo. Free gives a light agent quota in Devin Desktop, Cognition's own models and unlimited Tab autocomplete. Once the quota runs out you wait for a reset - on Free you cannot buy extra usage. Reviewing public PRs at devinreview.com is free and needs no account.
Devin pricing
2026-09-26
Limits
Daily and weekly quota; Enterprise - ACU
Pro and a full Teams seat get a daily and a weekly quota, Max only a weekly one. The price of a message depends on the model, the size and the complexity of the task; the vendor does not publish exact quota numbers. Beyond the quota, work runs on credits at API prices. Enterprise counts ACU: agent actions plus machine time. A sleeping session spends nothing, and Devin falls asleep by default after 30 minutes of idling.
Devin docs: self-serve plans
2026-09-26
● limits expire
All vendors in 2026 rebuilt billing around token credits; limit figures change every week. Each line above has a verification date - do not trust figures without a date, including ours.
6.0
Risks: privacy and control
Code privacy
No code training by default
By default, Cognition does not train models on customer data and code; SOC 2 Type II since September 2024. Enterprise: deployment in the customer's VPC, SSO, roles, IP lists. Output code belongs to the customer.
Devin docs: security
2026-07-06
Open source
No, the core is closed
There is no public product repository: github.com/CognitionAI has only supporting tools (blockdiff, deepwiki, SWE-bench results).
GitHub CognitionAI
2026-07-06
Company
Cognition: a round of more than $2B at a $48B valuation (Sep 2026)
Cognition raised more than $2B on Sep 8, 2026, in a round led by Andreessen Horowitz and Accel. Annualized revenue grew from $492M in May to $1B on Sep 25, 2026. In July 2025 Cognition bought Windsurf, and since Jun 2, 2026 it is called Devin Desktop. Customers include NVIDIA, Citi, Mercedes-Benz and GE Aerospace.
Cognition: Series E
2026-09-26
7.0
Devin or Claude Code - cloud versus terminal
Devin - a cloud fleet: assign a batch of tasks and accept PRs. Claude Code - terminal-first: depth in your own repo here and now. The choice is "where the work lives," not "who is smarter."
Devin
Claude CodeCardAA Index (best pair, Sep 25)
fleet of parallel cloud sessions; batch delegation
Strength
depth in a large codebase, careful edits
Free $0 without the cloud; Pro $20/mo; a task in AA Index $7.90 vs $13.04 for Claude Code with Opus 5.5
Economics
from $17/mo; limits in 5-hour windows + weekly
Cloud, Desktop (formerly Windsurf), CLI, Slack/Teams, tickets, API
Surfaces
14: CLI, IDE, desktop, web, iOS, CI, Slack, SDK
Fusion: Fable 5.1 or GPT-6 Astra paired with its own SWE-2; in CLI and Desktop the choice goes all the way to open source
Models
Claude only (API/Bedrock/Vertex/Foundry)
core is closed, no product repository
Openness
no, proprietary CLI
- →Devin is a second layer: a fleet for delegable batches on top of a terminal agent for daily work.
Other paths
Devin or Codexcloud tasks come with the ChatGPT subscription that many already have
Devin or Windsurftheir own IDE, now Devin Desktop: if you need an editor, not a fleet
Devin or Copilot Coding Agentdelegation directly in GitHub, if the team lives there
Frequently asked questions
Devin AI - what is it in plain words?
A cloud AI agent from Cognition. You set a task from the browser, Slack or a ticket, Devin works on its own in its cloud machine and returns a finished PR for review. It is an executor of tasks, not a suggester in your editor.
Devin or Claude Code - which to choose?
Different jobs. Devin takes a batch of tasks from Slack or tickets and returns PRs for review, leaving the developer's machine free. Claude Code works in the terminal right in your repository, here and now. In the Artificial Analysis index as of Sep 25, 2026 the best Claude Code pair scored 66 and Devin Fusion 61.7. A task with Devin is cheaper, though: $7.90 against $13.04. Devin is handy to keep as a second loop for batches you delegate.
How much does Devin cost?
Free at $0 gives a light quota in Devin Desktop without cloud Devin. Pro at $20 a month opens cloud Devin and the premium models, Max at $200 gives noticeably more quota. Teams costs from $80 a month, a full seat is $40 and counts toward that minimum. The price of a message depends on the model and the complexity of the task; the vendor does not publish exact quota numbers (prices checked Sep 26, 2026).
Is Devin free?
Partly. Free at $0 gives a light quota in Devin Desktop and unlimited autocomplete, but there is no cloud Devin on it. The cloud starts with Pro at $20 a month. Reviewing public PRs at devinreview.com is free and needs no account.
What is Devin Fusion?
A mode where two models run a task. The lead one, Fable 5.1 for example, plans and checks, while the cheap helper SWE-2 writes the code and runs the tests. In the Artificial Analysis index such a pair scored 61.7 at $7.90 per task, almost like Claude Code with Fable 5.1 (62.2 at $12.39). Fusion is in cloud Devin since Jun 29, 2026, and in Desktop and the CLI since Sep 11, 2026, on paid plans only.
What is ACU in Devin?
Agent Compute Unit - the unit of consumption in Enterprise contracts. It counts agent actions and cloud machine time. On self-serve plans quota and credits work instead of ACU, and at face value a credit equals an ACU. The "1 ACU = 15 minutes" formula from older reviews is not in the current documentation.
What is correct: Devin, devin, or devin ai?
The canonical name is Devin, a Cognition product. Do not confuse it with namesakes in npm: the official CLI is installed by a script from cli.devin.ai.
What happened to Windsurf after Cognition acquired it?
Cognition bought Windsurf in July 2025 and on Jun 2, 2026 renamed it Devin Desktop. It is the same IDE, now with the Devin Local agent and cloud Devin sessions inside. If you need an editor rather than a cloud fleet, see the Windsurf card.
Watch and read on the topic
Entity · devin
Facts · 13, each with a source and date
Card editorial review · 2026-09-26
Facts verified · 2026-09-26
Where to go next
Rating updates
New cards, price and limit changes, ranking shifts - by email, no spam.