This review is built from what could be checked on November 6, 2026: Devin's and Cognition's own pages (pricing, documentation, the model price list, changelogs, blog posts, terms, privacy policy and status page); a month-long test by Answer.AI; an analysis of the launch demo by the Internet of Bugs YouTube channel; prompt-injection research by Embrace The Red; a code-review benchmark run by Tenki, which sells a competing reviewer; Hacker News threads; reporting by TechCrunch, Built In, Dealroom, Tech Funding News and Channel Insider; and rivals' pricing pages. Cognition's claims about itself are labelled as its own.
What Devin is
Devin is the AI coding agent of Cognition AI, Inc., a San Francisco company that introduced Devin on March 12, 2024 as "the first AI software engineer". In 2026 Devin is less one product than a family under one account:
- Devin Cloud, the original agent. You give it a task in the web app, Slack, Microsoft Teams, Linear, Jira or through the API; it works on its own Linux, Windows or macOS virtual machine, clicks through what it built and opens a pull request with a screen recording.
- Devin Desktop, the editor Cognition got when it bought Windsurf in July 2025, renamed on June 2, 2026. Its Agent Command Center shows local and cloud agents on one board and also runs outside agents such as Codex, Claude Agent and OpenCode.
- Devin CLI, a local agent for the terminal that can hand a task to a cloud Devin and pull the finished branch back.
- Devin Review, a pull-request reviewer for code written by people or by agents.
Cognition is heavily funded. It raised more than US$2 billion at a US$48 billion valuation on September 8, 2026, led by Andreessen Horowitz and Accel, as Dealroom and Tech Funding News also reported, and says it passed US$1 billion in annualized revenue on September 25. Tech Funding News and Channel Insider, citing The Information, reported that its cash burn could reach about US$800 million this year.
What changed in October and November 2026
- SWE-2 Flash (November 4). Cognition's new small coding model is built on Inkling-Small from Thinking Machines, which Cognition calls a fully American base, and comes in Medium, High and Max effort. Cognition says it costs one-fifth as much as SWE-2 and runs three times as fast, and scores 24.0% on version 2.0 of FrontierCode, Cognition's own benchmark, against 23.8% for SWE-2. It runs only in Devin Desktop and the CLI. Cognition is not yet releasing the weights publicly, citing the risk of misuse in cyber attacks.
- Devin Review 2.0 (November 5). Cognition says the cost of reviewing a pull request fell 82% since August, to US$0.25 to US$0.75 for a typical review, while bugs found per PR rose from 1.22 to 1.32 and median time fell from 139 to 94 seconds. New are incremental reviews of each push, an Ultra effort level and custom reviewers in beta.
- Kimi K3.1 Pro. Moonshot's Kimi K3.1 Pro appears in Devin's model price list, dated November 5, 2026, in four effort levels at US$2 per million input tokens and US$10 per million output. No announcement post was found.
- Earlier this autumn: SWE-2 on September 10; sign-in with ChatGPT on September 29, which lets a ChatGPT Plus or Pro plan pay for GPT models inside Devin; Devin Oncall on October 15; Devin Security 2.0 on October 26, with more than US$100 million in credits offered to enterprises; and hierarchical spending budgets on November 2.
SWE-2 is post-trained from Moonshot's Kimi K3, a 2.8-trillion-parameter open model. Cognition says it scores 50.0% on FrontierCode 1.1 Main, within a point of Claude Fable 5.1 at 64% lower cost. Every one of these numbers comes from Cognition, and FrontierCode is Cognition's own test.
How Devin is priced
Cognition rebuilt its self-serve pricing on April 14, 2026. The US$500-a-month Team plan and the pay-per-ACU Core plan went away, and Ask Devin, DeepWiki and Devin Review stopped being free. This is the lineup on the pricing page:
| Plan | Price | Who and what | Devin Cloud |
|---|
| Free | US$0 | One user; light quota, limited models, unlimited Tab completions | No |
| Pro | US$20 a month | One user; daily and weekly quota; frontier models; SWE-2 free for now | Up to 10 sessions at once |
| Max | US$200 a month | One user; larger weekly quota, no daily cap | Up to 10 sessions at once |
| Teams | From US$80 a month | Up to 200 users; full seats US$40 a month each, flex seats free on shared credits | Unlimited sessions |
| Enterprise | On request | Billed in ACUs; SSO, admin controls, dedicated deployment | Unlimited sessions |
Each paid plan has a usage allowance that refreshes daily and weekly; Max has only a weekly one. Past it you buy on-demand credits, prepaid and charged at the token price of the model you use; they never expire and can top up automatically. How much a quota covers is not published: the pricing page says only that most users will never hit their limits. On Teams, Devin Review and Automations always draw from the shared credit pool, and when credits run out, automations stop and Review falls back to a plain diff viewer.
Devin's usage docs explain what you pay for. Cloud sessions count the agent's actions plus virtual machine time; Windows sessions use about 9% more than Linux; a session sleeps after 5 minutes idle and costs nothing while asleep or waiting on you. Large enterprise customers can join Cognition's AI Productivity Guarantee: if Cognition's own estimate of the engineering hours Devin delivered falls short of what they consumed, it issues free credits, up to US$10 million.
The models you pick move the bill most. Rates for use beyond the quota, from Devin's model price list:
| Model in Devin | Per million input / output tokens, beyond the quota | Note |
|---|
| SWE-2 Flash | Free until December 4, 2026; then US$0.10 / US$0.50 | Desktop and CLI only |
| SWE-2 | Free in Desktop and CLI through November 30, 2026; then US$3 / US$15 | Not free in cloud sessions |
| Kimi K3.1 Pro | US$2 / US$10 | Moonshot's model, in four effort levels |
| GPT-6.1 Sol | US$2 / US$10 | Can be billed to a ChatGPT plan instead |
| Claude Opus 5.5 | US$4 / US$20 | Fast variants cost twice as much |
| Claude Fable 5.1 | US$10 / US$50 | Anthropic's top model |
The price list has more than 250 model and effort options. For SWE-2 Flash, Cognition's launch post says "Free through 11/30", four days earlier than the price list's December 4 at 18:00 UTC.
Refunds are tight. The terms updated June 30, 2026 call payments non-refundable unless the terms or the law say otherwise, and Devin's docs offer a refund only before a subscription is used. There is no free trial of a paid plan. Windsurf subscribers who had a US$15 Pro plan keep that price after the switch to quotas in March 2026.
What public tests and users report
Third-party tests.
- Answer.AI, January 2025. Hamel Husain, Isaac Flath and Johno Whitaker gave Devin 20 tasks over a month: 3 successes, 14 failures, 3 inconclusive. Devin handled glue code, such as pulling a Notion database into Google Sheets, but produced overcomplicated code for an API integration, got stuck parsing HTML in a scraping job and spent more than a day on a deployment Railway does not support. Whitaker summed it up: "Tasks it can do are those that are so small and well-defined that I may as well do them myself, faster, my way." This was the 2024 Devin.
- Internet of Bugs, April 2024. A frame-by-frame video of the launch demo argues Devin did not complete the Upwork job it was shown doing and took at least six hours on work the author redid in about 36 minutes.
- Embrace The Red, August 2025. Security researcher Johann Rehberger showed that instructions planted in a GitHub issue made Devin download and run malware, and later posts showed it leaking secrets and opening ports to the internet. He reported the flaws in April 2025 and said Cognition had not given a fix timeline 120 days later. Devin now has security profiles that restrict network, MCP and git access.
A competitor's test. Tenki, which sells its own reviewer, ran seven tools on 50 pull requests holding 122 real bugs in April 2026. Devin caught 44 (36.1%) at 47.3% precision, an F1 score of 40.9, second to Tenki's 41.7 and ahead of Cursor (39.4), CodeRabbit (26.7), Greptile (22.1) and GitHub Copilot (21.4). That run predates Devin Review 2.0.
Cognition's own numbers. Cognition reported in November 2025 that 67% of Devin's pull requests were merged, up from 34% a year earlier, and described Devin as senior-level at understanding a codebase but junior at execution. Its own docs say Devin is most likely to succeed on tasks you could do in three hours or less, with clear success criteria.
User reports. In the Hacker News thread on SWE-2 in September 2026, opinions split. Critics called the CLI buggy, said the web and desktop clients run different harnesses of uneven quality, and objected that SWE-2 can only be used inside Devin. Supporters, including a heavy team user, called Devin Cloud one of the best team-managed cloud agents for its sandboxed testing and API, and one Pro subscriber found SWE-2 Max clearly better than Kimi K3. A separate commenter, who builds a competing product, wrote that Devin Automations has solid UX "but can get very expensive to run".
Reliability. Devin's status page lists 25 incidents from May 13 to November 4, 2026, including Devin Cloud being unavailable on August 13 and the web app going down on September 24.
Privacy, security and where your code runs
Training is on by default. Section 3.3.1 of the terms lets Cognition use customer data to train models and improve its services; a paid plan can opt out on the Data Controls page, after which Cognition says the data is not used and zero data retention is switched on with its model providers. On Teams only an admin can opt out, the Free plan cannot, and Enterprise customers are promised no training without written consent. Outputs belong to you.
Cognition says it has held SOC 2 Type II since September 2024 and works to ISO 27001, and the Windsurf extensions received FedRAMP High authorization in March 2025. The deployment docs say each session runs on its own isolated machine in Cognition's cloud on AWS. Enterprises can move Devin's workspace into a single-tenant network reached over AWS PrivateLink, or run sessions on their own machines with Outposts. Devin's "brain", the service that powers its reasoning, always stays in Cognition's cloud. The privacy policy, updated October 26, 2026, says Cognition shares identifiers with advertising partners in ways some US states treat as a sale, with an opt-out page.
Who should not use Devin
- Anyone who needs a fixed monthly cost. Quotas are not published in tokens or tasks, and use beyond them is billed per token.
- Teams that cannot write tight specs. Devin's own guidance asks for clear success criteria, and Answer.AI could not predict which tasks would succeed.
- Free-plan users with private code they do not want used for training. Only paid plans can opt out.
- Anyone pointing an unattended agent at untrusted issues or web pages without a network allowlist. The 2025 prompt-injection research shows what can happen.
- Developers who only want an assistant in the editor. Cheaper tools cover that, and Devin's strengths are in the cloud agent.
Alternatives
| Option | What it is | Entry price, checked November 6, 2026 | Better choice when |
|---|
| Devin | Cloud agent, desktop editor, CLI and PR review in one plan | Free; Pro US$20 a month | You want to assign tickets to an agent that tests on its own machine |
| Claude Code | Anthropic's coding agent, in every paid Claude plan | Claude Pro US$20 a month (US$17 billed yearly) | You drive the agent yourself and already use Claude |
| OpenAI Codex | OpenAI's agent on the web, in the CLI, IDE and iOS, with cloud code review | Free and Go (US$8) with GPT-6 Luna; Plus US$20 | You already pay for ChatGPT |
| Cursor | AI editor with cloud agents and automations; bought by SpaceX in August 2026, per Tech Funding News and Channel Insider | Premium US$30 a month for cloud agents; Plus US$8 runs only Grok models | You live in an editor and want agents as an add-on |
| GitHub Copilot | GitHub's agent, cloud tasks and code review, plus Claude Code and Codex as agents | Pro US$10 a month | Your code is on GitHub and you want the cheapest cloud agent |
| Factory | Droid agents on desktop, CLI and in the cloud | Pro US$20 a month; Teams US$60 plus US$40 a seat | You want a similar agent with a different harness |
Several of these bundle an agent into a plan you may already have. Devin's edge is a managed cloud computer per task that tests its own work, hand-offs from Slack, Linear and Jira, and enterprise deployment options; its weak spots against them are unpublished quotas and thin outside testing. On GenGrades, T3 Code is a free app that runs the agents you already pay for side by side, much like Devin Desktop's command center; ZCode is Zhipu's free desktop coding agent on its GLM models; Command Code is another terminal agent; and OpenCode Zen sells models for the open-source OpenCode agent. If you want an agent for research or slides rather than code, Manus is the closer fit.
Before committing a team, run the same ticket through Devin and one alternative and compare the merged result and the cost.