Devin arrived with enormous hype as the first “AI software engineer” capable of completing entire engineering tasks autonomously. The benchmarks Cognition published were impressive. The actual product, when it shipped, was more complicated than the demo suggested. That’s not a knock on Devin specifically; the gap between impressive demos and daily-driver reliability is a universal challenge in this category. But it’s worth being clear-eyed about what Devin actually is today.

Correction, September 2026: this page had Devin’s entry price wrong

This review carried $500/month as Devin’s starting price. That has not been true for months. Devin’s published plans are Free ($0), Pro ($20/mo), Teams ($80/mo base plus $40 per full seat) and Max ($200/mo), with a custom Enterprise tier above that. The $500 figure was the old standalone cloud-Devin price and it is gone.

We’re stating that plainly rather than quietly editing it, because the error changed the advice. Several sections below argued that Devin was priced out of reach for anyone without an engineering budget. That specific objection no longer holds. What follows has been rewritten where the price was doing the work in the argument, and left alone where it wasn’t.

The rating stays at 2, and the reason is worth separating from the price. Devin was never a bad fit for non-technical founders because it cost too much. It was a bad fit because you cannot write a ticket precise enough for it to work from, and you cannot tell whether the pull request it hands back is correct. A $20 seat does not fix either of those. It just removes the excuse you were using instead of the real reason.

New in September 2026: SWE-2 lands, and the $20 plan gets the interesting model

On September 10, Cognition shipped SWE-2, its own coding model, and made it free for every Pro, Max and Teams subscriber for a month — meaning it does not draw down your usage quota at all. Cognition’s growth lead put it as “unlimited access to SWE-2 for all pro+ plans (starting at $20/month).” It’s live in Devin Desktop and the Devin CLI, rolling out to Devin Web and Fusion. There is no API and no open weights: SWE-2 only runs inside Devin.

The model is post-trained with reinforcement learning from Moonshot AI’s Kimi K3, a 2.8-trillion-parameter open model. Cognition’s published numbers put it at 50.0% on FrontierCode 1.1 Main against 50.9% for Claude Fable 5.1 and 53.3% for GPT-6 Astra, at a claimed 64% lower cost than Fable. Two caveats you should hold onto: FrontierCode is Cognition’s own benchmark, and every cost figure here is the vendor’s, unaudited. On Terminal-Bench 4 — long, messy, multi-step terminal work — SWE-2 scores 27.3% against Fable’s 55.8%. That is not a rounding error, and it’s the honest ceiling on what this model does well.

The more useful number is behavioural. Cognition trained SWE-2 with a cost penalty: the model gets docked for burning money. The previous generation averaged 127 steps per task and didn’t make its first real edit until step 48. SWE-2 on medium effort averages 53 steps and starts editing at step 18 — a better score with 58% fewer turns and 81% lower average cost. Anyone who has watched an agent read an entire repository before touching a file will recognise what was fixed.

What this means if you’re not an engineer. Not much directly, and that’s the point worth sitting with. The story here isn’t “Devin got cheap enough for you.” It’s that tool vendors are now training house models specifically so the good option can be the free one — the same move Replit made with Luna, and the pattern we’ve written up in Why vibe coding tools are building their own models. The savings land in your quota, not on the pricing page. If you do run a Devin seat, the free month ends around mid-October; SWE-1.5 and SWE-1.7 both stayed free on Pro after their promos ended, which is a decent prior and not a promise.

The same day brought Devin Voice, a full-duplex voice mode running on GPT-Live plus SWE-2 — you talk, it works in the background, and either of you can interrupt. Treat it as a demo-stage feature until people have run it on real work for a few weeks.

New in September 2026: the round closed bigger than reported — $2B at $48B

The raise we flagged on September 2 as “roughly $1 billion at ~$47 billion” closed on September 8 at more than $2 billion at a $48 billion post-money valuation, led by a16z and Accel. Run-rate revenue is reported at roughly $900 million, up from $492 million at the May round. Named customers now include NVIDIA, GE Aerospace, Citi, Mercedes-Benz and Modal.

Worth connecting to the section above: a company that just raised $2 billion can afford to give a near-frontier coding model away for a month to buy distribution. The free SWE-2 promo is customer acquisition and a training-data flywheel, not generosity, and reading it that way will keep you from being surprised when the terms change.

New in September 2026: Devin moved its code review off Opus 5 the day Fable 5.1 shipped

When Anthropic released Claude Fable 5.1 on September 1, Cognition’s Walden Yan was quoted in the launch materials confirming Devin was moving traffic that day: “with the new cache read pricing a Fable-class model is finally economical for the workloads we’d kept on Opus, starting with code review.” Anthropic’s sticker price didn’t change — the cache read price dropped 75%, from $1.00 to $0.25 per million tokens, which cuts the real cost of heavily agentic work by up to 45%.

This is a small item with a useful lesson attached. Devin’s underlying model changed on a Tuesday and the price you pay did not. That cuts both ways: you got a capability upgrade you weren’t billed for, and you have no visibility into it. If Devin’s behaviour on your codebase shifted in early September — different review comments, different pacing — this is the likely cause, and nothing in your account settings would have told you.

It also sharpens the economics question below. Cognition’s margin on Devin improved and the price on the page did not — which is the same mechanic, running in the vendor’s favour this time, that SWE-2 now runs in yours.

New in September 2026: Cognition raising again at $47B, fourteen weeks after the last round

Superseded — the round closed on September 8 at more than $2B and a $48B valuation. See above. The reporting below is kept because the read on why appetite spiked still holds.

Bloomberg reported on September 2 that Cognition is closing a round of roughly $1 billion at a ~$47 billion valuation. That is not a typo and it is not the $40B figure that circulated in August — the number moved. It would nearly double the $26 billion post-money the company closed in late May, in about fourteen weeks. Investor interest reportedly ran to nearly $10 billion against a $1 billion raise. Cognition declined to comment. Source: Bloomberg via Tech Funding News.

The revenue is doing some of the work here. Annualised run-rate has gone from $492 million at the May round to over $900 million — roughly doubling in the same fourteen weeks. Whatever you make of the valuation, that is not a company waiting for product-market fit.

The other half of the story is who’s left. SpaceX’s $60 billion acquisition of Anysphere — Cursor’s parent — legally closed on August 14, and Bloomberg’s sources point to that close as one reason appetite for Cognition spiked: it is now one of very few large, independent bets available in AI-native software development. Meanwhile Lovable raised at $13.3 billion in August and Replit at $9 billion in March, both aimed at the non-developer end of the market Devin has never targeted.

Nothing here changes the recommendation on this page. Rating held at 2. A round this size confirms what the May round already told you: Cognition is building for engineering orgs and has the capital to keep doing it for years. It did make Devin cheaper at the entry point — but cheaper access to a tool you can’t evaluate the output of isn’t an improvement in your position.

New in August 2026: Devin moves into Slack channels

On August 20, Devin became one of four founding partner agents in Slack Code, alongside Claude Code, GitHub Copilot, and Vercel’s agent. Cognition ran the launch demo: someone reports a broken feature in an engineering channel, Devin acknowledges it in the thread, investigates, tags the code owner, and opens a pull request — using Chrome and DevTools on its own machine to verify the feature actually works, then posting screenshots and a recorded demo as proof. A designer dropped a Figma file into the channel mid-task and Devin incorporated it without stopping.

That verification loop is the most substantive part of this update, and it plays to a real advantage of cloud agents over ones running on a laptop: they can produce an auditable record that the work is correct. Security-wise, Devin runs in isolated sandboxes with what Cognition calls minimum viable access, including an optional mode with no internet access at all.

Cognition also offered a velocity number: internally, merged PR count up 10x over a few months while headcount rose about 40%. Treat that with the scepticism any vendor-supplied metric deserves — merged PRs reward volume, not value — but the direction is consistent with what teams running Devin report.

None of this changes the rating of 2. Devin still needs someone who can evaluate its output, and Slack Code makes that person’s job easier rather than optional. Cognition’s Jeff Wang did predict that the human approval gate before merge “maybe in the next year is just going to go through automatically.” If you run Devin, that’s the setting to keep an eye on. Source: VentureBeat, Unite.AI.

New in June 2026: Windsurf becomes Devin Desktop, and the stack consolidates under one name

On June 2, Cognition folded its IDE into the Devin brand: Windsurf is now Devin Desktop. What used to be two products with two names is now one product line: Devin Desktop for interactive, human-in-the-loop work in the editor, and cloud Devin for background, ticket-driven runs. The local agent that powered Windsurf (Cascade) is being replaced by Devin Local, a Rust rewrite Cognition says is ~30% more token-efficient, and the desktop app now opens on the Agent Command Center by default. Cascade hits end-of-life on July 1, 2026. Source: Devin blog.

For non-technical founders the takeaway is unchanged: it’s all developer-shaped. Devin Desktop (formerly Windsurf) is the editor, cloud Devin is the async agent, and both now sit inside the same Free/$20/$200 plan ladder. What changed is the story. Cognition now has a single, legible product ladder from free IDE to autonomous cloud engineer, all branded Devin. See our Windsurf review for the desktop side of that ladder.

New in May 2026: Cognition closes $1B at a $26B valuation

The funding talks we flagged in April have closed, and then some. On May 27, Cognition confirmed it raised more than $1 billion in a Series D led by Lux Capital, General Catalyst, and 8VC, at a $25B pre-money / $26B post-money valuation. That’s up from a $10.2B valuation just eight months ago, roughly 2.5x in under a year. Cognition says Devin is now generating around $492M in annualized revenue run-rate, with enterprise usage growing ~50% month-over-month for six straight months, and names customers including Mercedes-Benz, NASA, Goldman Sachs, and Santander. The company also claims Devin now writes the large majority of Cognition’s own code, a self-referential proof point that’s either genuinely impressive or a marketing flourish, depending on how much you trust the framing.

For non-technical founders, the practical takeaway hasn’t changed: Devin is still the wrong starting point. (The price prediction in this paragraph aged badly — the entry point came down to $20 rather than staying put. The recommendation didn’t depend on it.) What the raise does signal is staying power. Cognition now owns both the IDE (Windsurf) and the async agent (Devin), has a deep war chest, and is clearly aiming at enterprise engineering-org budgets rather than the indie-builder market that tools like Lovable and Bolt serve. If you’re a technical founder wondering whether Devin will still be around in two years, this round answers that. If you’re non-technical, it mostly confirms Devin isn’t being built for you.

New in April 2026: Devin for Terminal ships inside Windsurf

On April 28, Cognition shipped Devin for Terminal as part of Windsurf 2.1.29. Every Windsurf user now has Devin available as a CLI agent included with their existing subscription, with no additional Devin license required. The terminal agent runs locally with full access to your codebase, tools, and environment, and hands sessions off cleanly to the cloud Devin instance when you want a longer-running task to keep working in the background. Cognition reports the new agent is roughly 30% more token-efficient than the existing Cascade agent and supports a model picker that includes Claude Opus 4.7, GPT-5.5, and Cognition’s own SWE-1.6.

The strategic read: this is the first time Cognition has bundled the full Devin experience inside the Windsurf subscription rather than the then-$500/mo standalone product. For non-technical founders, that doesn’t really change the recommendation. Devin is still the wrong starting point, and Windsurf is still developer-shaped. But for technical founders or PMs already paying for Windsurf, you now get an async, autonomous coding agent without paying separately for it. That’s a meaningful unlock for teams that wanted Devin’s behavior but couldn’t justify the line item.

New in April 2026: Cognition in funding talks at a $25B valuation

Bloomberg reported on April 23 that Cognition, the company behind Devin, is in early talks to raise hundreds of millions of dollars at a $25 billion valuation, up from the $10.2B it hit in September 2025 after its $400M raise and more than double the valuation of roughly a year ago. Cognition also acquired Windsurf (the VS Code fork) in July 2025, and per subsequent reporting the combined ARR has roughly doubled since that deal, with Windsurf alone now contributing around $82M ARR and enterprise revenue compounding quarter-over-quarter. The story matters for Devin users in two ways. First, Cognition has the runway to keep pushing the “autonomous engineer” thesis even if it doesn’t monetize as fast as a subscription tool like Cursor, and the pricing is unlikely to drop. Second, the company now owns a full stack from IDE (Windsurf) to async agent (Devin), and a consolidated product story is starting to emerge: Windsurf for interactive, human-in-the-loop work; Devin for background, ticket-driven runs. If that integration lands cleanly, the case for Devin inside an engineering team gets stronger; if it doesn’t, you end up paying the then-$500/mo standalone price for a tool that lives awkwardly next to the editor your engineers actually use. (That price came down to $20 later in 2026 — see the correction at the top of this page.)

What Devin does

Devin is an agentic AI engineer that takes a task, spins up a development environment, writes code, runs tests, iterates on failures, and produces a finished result, theoretically without requiring you to hold its hand at every step. You describe what you need in natural language (or a GitHub issue), Devin works on it asynchronously, and you review the pull request when it’s done.

The tasks where Devin works well are well-defined, bounded engineering tasks: adding a feature with a clear specification, writing tests for existing code, fixing a specific bug, migrating code to a new library. Think of it as an async contractor you can throw clearly scoped tickets at.

The autonomy caveat

“Autonomous” is doing a lot of heavy lifting in Devin’s marketing. In practice, Devin works autonomously in the way that a junior engineer works autonomously, which is to say it produces output that needs to be reviewed carefully before merging. On complex tasks with ambiguous requirements, it will make architectural decisions that may not match your codebase’s conventions. It can go down wrong paths for a while before self-correcting or getting stuck.

This isn’t a disqualifier; it’s just the accurate description. The value proposition is that Devin can run in the background while your engineering team focuses on higher-leverage work. A task that would have taken an engineer a few hours can be handed to Devin and reviewed 20 minutes later. At scale and with the right tasks, that’s valuable.

What it actually costs

Plan	Price	What you get
Free	$0	Light daily and weekly quota, limited models, unlimited Tab completions
Pro	$20/mo	Bigger quota, frontier OpenAI/Anthropic/Google models, cloud agents, free use of Cognition’s own SWE models
Teams	$80/mo base + $40 per full seat	Shared billing, admin, collaboration
Max	$200/mo	Much larger weekly quota, no daily cap, unlimited concurrent sessions
Enterprise	Custom	SSO, VPC deployment, usage-based billing

Two things about this table matter more than the numbers in it.

The subscription buys an allowance, not unlimited runtime. Each plan carries a daily and weekly token budget. Past it you buy more at API rates that Devin does not publish on the pricing page. An agent that gets stuck looping on a hard problem can eat a meaningful chunk of a week’s quota producing nothing, and there’s no way to forecast that from the page.

Which model you pick determines your real bill. Routing every task to Fable 5.1 or GPT-6 Astra burns quota fast. Cognition’s own SWE models don’t count against it at all. That’s the whole reason SWE-2’s launch matters commercially: the free model got good enough that most people won’t reach for the metered ones.

Teams is the tier people under-budget. It is not $80 flat. Five developers is $80 plus $200, so $280 a month.

Who should not use this

The non-coder rating is 2, and it really means it. Without technical knowledge, you cannot write specifications clear enough for Devin to work from, you cannot evaluate whether the output is correct, and you cannot integrate the results into a real product. Devin is a force multiplier for engineers, not a replacement for having engineering capability.

Verdict

Devin is genuinely impressive technology that was productized faster than the reliability curve warranted, and it has spent 2026 catching up. It’s most valuable as a background task processor for engineering teams with the oversight capacity to work with its output.

The price is no longer the reason to say no. At $20 a month with a free tier underneath it and a near-frontier model included, the cost objection that dominated this review for most of the year has collapsed. What replaced it is a cleaner question: do you have someone who can tell a good pull request from a bad one? If yes, a $20 seat is a cheap experiment and you should run it on three real tickets before deciding anything. If no, the price was never what was stopping you.

Was this helpful?