Cognition launched Devin for Databricks Migrations, built with Databricks, to help companies move data from older platforms onto the Databricks platform. Devin converts the old code, tests it in the customer's workspace, fixes failures, and checks the data against the new copy before switching over. A Fortune 500 energy company reportedly moved 142 objects in eight days.
Devin converts legacy data warehouse code and tests it in the customer's Databricks workspace.
It fixes failing parts and checks every source against its Databricks copy before cutover.
Devin can run hundreds of tasks in parallel, which speeds up large migrations.
One Fortune 500 energy company moved a 142-object platform to Unity Catalog in eight days.
Cognition, the company behind Devin, launched Devin Security Threats, a managed list of newly disclosed software vulnerabilities that connects to its Devin scanning tool. Users can scan their code for a specific threat with one click and have Devin propose fixes as pull requests. The launch covers two new vulnerabilities in widely used foundational software disclosed the same week.
Devin Security Threats is a managed index of newly disclosed vulnerabilities.
A threat describes what could go wrong, which systems are affected, how an attacker could use it, and how to fix it.
Users can pick repositories to scan, and Auto-triage can scan them automatically when new threats appear.
Devin ranks findings against the threat and opens pull requests that teams can review.
Devin, a coding agent from Cognition, now supports Gemini 4 Neon in its desktop app and command-line tool. Cognition says it is the best-scoring Google model on its FrontierCode 2.0 engineering test.
Gemini 4 Neon scored 46.2% on FrontierCode 2.0, the best result so far from a Google model.
That is about one point behind Claude Opus 5.5, which scored 47.3%.
Gemini 4 Neon costs about $4.38 per task, roughly a third of Claude Opus 5.5's $12.82.
Cognition has launched Devin Bots, persistent AI engineering agents that own a goal and direct several Devin coding agents in parallel. They keep their own memory across sessions and bring only the decisions that need a human to the user.
Devin Bots are persistent agents with their own memory and tools.
A Bot can lead multiple Devins working in parallel on a large project or ongoing goal.
It investigates, delegates work, evaluates results, and decides what to pursue next.
Examples include maintaining an open-source repo, keeping Android at parity with iOS, managing releases, and optimizing server inference.
Devin Apps are private web apps that run on Devin's cloud computer. Teams describe the tool they need, Devin builds it, and teammates open it from a link. The apps can use existing integrations and codebases.
Apps require sign-in and are never public.
Sharing can be limited to chosen people, the whole organization, or the whole enterprise.
Edits create new versions at the same URL, and if a new version fails, the previous one keeps running.
When unused, the machine sleeps and the first load can take a few seconds; billing depends on awake time.
Cognition, the maker of Devin, has updated its Teams plan. Every team member can now use Devin Desktop without paying extra. A free mode gives users free sessions powered by its own coding model, SWE-2.
Every team member can use Devin Desktop at no extra cost.
Each user gets 2 free sessions at the same time, and can pay for more.
Admins can connect a company login system (SSO) and set usage limits for each user.
Teams customers can get a standard health data protection agreement (BAA) without a large enterprise contract.
Cognition added the Gemini 4 Hydrogen model to its AI development tool Devin, in both the desktop app and the command-line tool (CLI). It scored 38.2% on FrontierCode 2.0, a test of real engineering tasks.
Gemini 4 Hydrogen scored 38.2% on FrontierCode 2.0.
On the same test, GPT-6.1 Sol scored 39.0% and Gemini 4 Carbon scored 40.0%.
Gemini 4 Hydrogen costs $2.40 per task, about 40% less than Gemini 4 Carbon's $4.04.
At medium reasoning effort, it beats Gemini 4 Argon's best score at under a seventh of the cost per task.
Cognition's AI development tool Devin now lets users run Gemini 4 Carbon in the Devin Desktop app and Devin CLI, a command-line tool. The company says the model scored higher than its predecessor on its own test of real engineering tasks.
Gemini 4 Carbon scored 40.0% on the FrontierCode 2.0 test.
That is well above the 26.3% of its predecessor, Gemini 4 Argon.
It also edges out GPT-6.1 Sol's 39.0% on the same test.
At low reasoning effort, it already beats Argon's best score.
Claude Fable 5.2 is now available in Devin, an AI developer tool. It scored 49.1% on FrontierCode 2.0, a test of real engineering work, taking the top spot.
FrontierCode 2.0 measures code quality and whether changes can be merged into real projects.
Fable 5.2 improved strongly over Fable 5.1's 32.4% and beat Opus 5.5's 45.5%.
At low reasoning effort (how much effort the AI spends on an answer), Fable 5.2 already beat Fable 5.1's best score, at less than a third of the cost per task.
Cognition has launched Devin Review 2.0, an AI code review tool. Since August, the cost of reviewing a pull request has dropped 82%, so it is now cheap enough to run on every one. Repeat reviews now check only the changed parts, lowering costs further.
A typical review now costs about $0.25 to $0.75.
Bugs found per pull request rose from 1.22 to 1.32.
Review speed improved 1.5 times, with median response time falling from 139 to 94 seconds.
Users can start a review on GitHub or GitLab with a '/devin review' comment.
Devin, the AI software engineer, launched hierarchical budgets that let enterprises manage AI spending to match how they are organized. Admins give ACU budgets (Devin's unit of usage) to department leaders, who then set and approve usage limits for their sub-teams.
Child budgets can sit under the company-wide budget, up to three levels deep.
Each budget can have its own monthly limit, a temporary bump for the current monthly cycle, and per-user limits.
Department heads see and manage only their own budgets, and requests for higher limits go to them.
Setting aside a dedicated budget for a project such as security hardening lets companies track that spending separately.
Devin released Devin Security 2.0, a security tool that scans code with frontier AI models to find vulnerabilities (security weaknesses), ranks them by real risk in your environment, and drafts code change requests to fix them.
It can scan with frontier security models such as Claude Mythos, and companies do not need separate access to those models.
It uses an approach called Agentic MapReduce, splitting code into batches investigated in parallel and then combining results to keep costs down.
Important findings are checked in an isolated environment against a running instance of the app.
It targets a world where AI is speeding up attacks, as with the recently disclosed remote code execution flaws (attacks that run commands on a server from outside) in PHP and Node.js.
Devin introduced three new enterprise cost-control features. Instead of putting the same usage cap on everyone, they allocate spending based on how efficiently each engineer uses Devin and tell less efficient users what to fix.
Efficiency Score gives each engineer a 0 to 100 rating based on recent Devin use.
Devin Coach flags mistakes such as vague requests or picking an expensive mode for a simple task before a prompt runs, and suggests fixes.
Engineers with good efficiency grades get their limits raised automatically as they near them, up to a ceiling the admin sets.
The score also credits investigation and planning work that leaves no PR (a code change request) by estimating the hours a human would have needed.
Devin made Devin Oncall, a tool that helps handle service outages, generally available. It investigates alerts and bug reports on its own to find root causes and helps teams diagnose and resolve problems in live services.
On SREGym, a benchmark built from real incidents, it achieved a 100% success rate with the fastest time to mitigation.
At Reducto, it caught warning signs and opened an incident before the company's own alerts fired, greatly reducing the impact.
At AngelList, it found the right root cause 3.5 times more often than the in-house tool and handled half of incidents end to end.
Some of the world's biggest companies now rely on it for thousands of live alerts each day.
Google's new Gemini 4 Argon model is now in Devin Desktop and the Devin CLI (command-line tool). Until October 27, it can be used at 75% below list price.
The price combines Google's 50% launch discount, running through December 1, with an extra 50% off in Devin through October 27.
On Devin's in-house coding benchmark FrontierCode 1.1, it scored 60.5%, edging past GPT-6.1 Sol (60.4%) and on par with GPT-6 Sol (60.7%).
It is the highest-scoring Gemini model on that leaderboard, four points ahead of Gemini 3.7 Flash.
It reads widely and reviews its own changes before finishing, and is strongest at bug fixes and refactoring (restructuring code) in backend and systems code.
Devin can now connect directly to the customer support tool Pylon. Support teams can hand investigations to Devin from Pylon, and engineers can pick up the first round of work Devin has already done.
When a ticket is flagged for engineering, Devin checks logs, reproduces the bug, finds the root cause and opens a fix PR (a code change request).
Progress syncs automatically in both directions between Pylon and Devin, so no one has to pass updates along by hand.
Beyond bugs, it can search the codebase to answer product questions and create screenshots and demo videos.
Devin is offering $15,000 in credits to new customers who use Pylon.
Claude Haiku 5.5 is now available in Devin. It outscored Sonnet 5 on Devin's in-house coding benchmark while costing about one-eighth as much per task.
It scored 58.4% on Devin's FrontierCode 1.1 benchmark, ahead of Sonnet 5.
It joins the helper (sidekick) models in Devin Fusion, which pairs a lead model with a helper model.
With Haiku 5.5 as the helper, Fusion keeps a top-tier score of 66.2 while cutting cost and delay.
Users can try it now in the Devin CLI (command-line tool) with Opus 5.5 as the lead model.