The latest AI news and updates of Cognition — Agent lab building Devin, a collaborative AI teammate for software engineering teams. Covering Cognition's latest product updates, company news, and launches from the past 90 days.
Cognition21h agoCognition introduced Fusion, a multi-model harness for Devin CLI and Desktop that pairs a frontier lead model for planning with a cost-effective sidekick for execution. This architecture maintains frontier performance while reducing costs by up to 39% across coding benchmarks. The lead model retains control, reviewing work and managing the session to ensure quality and efficiency.
CognitionSep 10Cognition launched Devin Voice, a conversational interface for its AI software engineer. Powered by GPT-Live and the SWE-2 model, the feature supports verbal interaction to explore ideas, test approaches, and hand off tasks. The interface includes microphone controls, audio silencing, and background navigation, with all conversation history saved directly to the active session.
CognitionSep 10Cognition launched SWE-2, its most advanced coding model, achieving 50.0% on FrontierCode 1.1 Main. By scaling reinforcement learning to multi-trillion parameters, SWE-2 matches Fable 5.1 performance at 64% lower cost. The model introduces support for configurable effort levels and is available today in Devin Desktop and CLI, with a one-month free period for Pro, Max, and Teams subscribers.
CognitionSep 9Cognition factored RSA-260, a 260-digit number, using a new GPU-optimized lattice siever built by its AI software engineer Devin. This implementation makes factoring 10x cheaper than the previous state of the art, costing approximately $400,000 in GPU time. The project demonstrates Devin’s ability to autonomously handle complex performance engineering and cluster management for large-scale computational research.
CognitionSep 8Cognition raised over $2B in Series E funding at a $48B valuation, led by Andreessen Horowitz and Accel. The company reports run-rate revenue growth from $492M in May to nearly $900M. Devin now features autonomous alert remediation via Slack and a Security Swarm for vulnerability triage, supporting enterprise engineering teams at NVIDIA, GE Aerospace, Citi, and Mercedes-Benz.
CognitionSep 3Cognition is adding OpenAI's GPT-6 Astra to Devin. On the FrontierCode 1.1 benchmark, Astra performs within 0.4 points of Fable 5 at a 64% lower cost. The model also sets a new SOTA on internal testing benchmarks, producing more comprehensive reports. Enterprise customers in OpenAI's Daybreak Program have access today, with broader rollout following in the coming days.
CognitionSep 1Devin, Cognition's AI software engineer, now runs on Fable 5.1, dropping task costs 54% thanks to Anthropic's 4x reduction in prompt-cache pricing (to $0.25/M cached tokens). Since 95%+ of Devin's tokens are cached, the pricing change has outsized impact. Cognition's Fusion multi-model harness hits the same FrontierCode 1.1 score for $1.43 per task — 47% cheaper than Fable 5.1 solo.
Devin DesktopAug 31Cognition updated the Devin CLI with a unified model picker. The interface integrates model selection, reasoning effort adjustment, and cost comparison into a single view. Accessible via the /model command, the update provides direct control over model behavior and efficiency for coding tasks.
CognitionAug 26Cognition overhauled the Devin webapp with a rebuilt chat renderer that reduces loading lag by 80% and speeds up long session openings by 55%. The update adds a customizable sidebar with compact views and grouping, a new ⌘K command menu for keyboard navigation, and native management for nested subagents. These changes improve performance and navigation for large-scale coding sessions.
CognitionAug 22Cognition is applying a 20% price reduction from OpenAI to its existing 70% discount for GPT-5.6 Sol, resulting in a 76% total discount off list prices. This pricing, effective through October 3, 2026, brings the model to $1.20 per 1M input tokens and $6 per 1M output tokens for users on Devin Desktop and Devin CLI.
CognitionAug 22Cognition reports that GE Aerospace is using Devin to build aviation software. One engineering team nearly doubled its output while developing systems for route optimization, fuel reduction, and anomaly monitoring. The deployment supports the company's goal to increase flight capacity without adding new aircraft, runways, or pilots.
CognitionAug 22Cognition introduced Slack Code in Devin, where the agent proactively creates dedicated code channels for complex tasks. When Devin identifies a task requiring more space than a thread, it automatically ports the session to a new channel. These channels feature a status chip, context bar, and artifacts tab to keep team work focused and transparent.
CognitionAug 13Cognition integrated Gemini 3.7 Flash into Devin Desktop and CLI. On the FrontierCode 1.1 benchmark, the model reaches Claude Sonnet 5-level performance at less than half the cost. It excels at tightly scoped refactors and matching repository conventions. A 50% discount on the model is available within Devin through August 27, 2026.
CognitionAug 13Cognition integrated MongoDB Atlas into Devin, enabling the AI agent to autonomously query, manage, and provision database clusters during software engineering tasks. This integration removes the need for manual database setup or static schema context, allowing the agent to interact directly with live data environments while building and testing applications.
CognitionAug 12Cognition integrated xAI's Grok 4.6 into Devin Desktop and the Devin CLI. On the FrontierCode 1.1 Extended benchmark, the model achieves a 61.3 score, surpassing GPT-5.6 Sol while trailing Claude Opus 5 and Claude Fable 5. The model demonstrates specific strengths in code exploration, root cause analysis, repo convention matching, and testing rigor.
CognitionAug 4Cognition updated Devin Fusion, achieving a 4% increase in intelligence and a 27% reduction in costs on the FrontierCode 1.1 benchmark. These gains result from refinements to both the model harness and the underlying models. Devin Fusion is available now in Devin Cloud for engineering teams.
CognitionAug 1Cognition now enables Devin to build and test native iOS apps using Devin Outposts. By running sessions on macOS environments powered by Namespace Devboxes, Devin gains access to Xcode, the iOS simulator, and Apple M5 chips. The agent autonomously builds, tests, and verifies iOS applications, sending screen recordings and test reports directly for review.
CognitionJul 30Cognition updated Devin to natively support GitHub Stacked PRs. The agent automatically decomposes large tasks into smaller, independently reviewable diffs, addresses review comments across the stack, and handles rebasing for downstream changes. This update maintains stack coherence and CI verification, allowing for focused reviews of complex code changes without manual bookkeeping.
CognitionJul 27Cognition added the open-source Kimi K3 model to Devin Desktop and CLI. On the FrontierCode 1.1 Extended benchmark, the model achieves a 58.2% score and 63.6% pass rate. It demonstrates strength in reproducing bugs and managing environments, though it occasionally deviates from stated specifications.
CognitionJul 27Cognition joined the NVIDIA-led Open Secure AI Alliance to support open-source models. The company is contributing research on evaluating model trustworthiness, specifically testing for propaganda, problematic requests, and insecure code generation. Cognition’s results indicate these risks are not inherent to open models and can be mitigated through careful post-training, arguing that open-source models remain essential for innovation.
CognitionJul 24Cognition integrated Anthropic’s Claude Opus 5 into Devin. On the FrontierCode 1.1 Extended benchmark, the model achieves a 63.6% score and 69.6% pass rate, approaching Claude Fable 5 performance at half the cost. The model demonstrates particular strength in complex debugging and root-cause analysis tasks within the agentic coding environment.
CognitionJul 21Cognition launched Devin Outposts, a feature that executes Devin Cloud sessions on user-controlled infrastructure. The agent’s reasoning loop remains in the cloud, while command execution, file edits, and repository access occur on local machines, VMs, or Kubernetes clusters. Deployment templates support platforms including Cloudflare, Modal, and NVIDIA Brev for private network and GPU-intensive workflows.
CognitionJul 21Cognition acquired TierZero, a startup building autonomous agents for incident response, alerts, and system reliability. This acquisition integrates TierZero’s operational capabilities into Devin Automations. The platform now extends beyond code generation to include autonomous monitoring and incident management, as Devin handles software maintenance tasks once code ships.
CognitionJul 18Cognition launched a live, public FrontierCode leaderboard to track how well AI models write production-quality code. The page ranks 17 models, including Grok 4.5 and Inkling, based on mergeability, pass rates, and rollout costs. It provides full methodology and interactive sample tasks, with automated detection to zero out runs that consult solution-bearing sources during evaluation.
CognitionJul 16Cognition launched Devin for Startups, providing early-stage companies with $65,000 in total value. The program includes $15,000 in immediate credits and $50,000 in matching grants, alongside unlimited team seats and exclusive founder events. Accepted startups gain access to Devin across its cloud, desktop, and CLI interfaces to accelerate software engineering workflows.
CognitionJul 15Cognition launched Devin into general availability for engineering teams at $500 per month with no seat limits. The release includes a Slack integration where tagging @Devin triggers issue investigation, codebase answers, and development tasks directly in the channel. The plan also provides access to the Devin IDE extension, API, and onboarding support.
CognitionJul 8Cognition launched SWE-1.7, its most capable model for agentic coding, built on a Kimi K2.7 base. The model achieves a 42.3% score on the FrontierCode 1.1 Main benchmark at a cost of $1.97 per task. It is available today in Devin across Web, Desktop, and CLI, served via Cerebras at 1,000 tokens per second.
CognitionJul 2Cognition launches a six-week Security Vulnerability Remediation Program to clear vulnerability backlogs. The service deploys Cognition engineers and the Devin AI agent to ingest reports from existing scanners, validate findings, and ship remediation pull requests. This program establishes continuous discovery and patching workflows, shifting security teams from reactive triage to proactive maintenance without pulling engineering time from product roadmaps.
CognitionJul 1Cognition makes Anthropic’s Claude Fable 5 model available across the Devin platform. The model integrates into Devin Cloud’s Ultra agent, which targets long-horizon tasks and debugging, as well as Devin Desktop and the Devin CLI. This update follows the global redeployment of Claude Fable 5.
CognitionJul 1Cognition launched Devin Security Swarm, a security scanning tool powered by a new Agentic MapReduce architecture. It identifies vulnerabilities across codebases, validates exploitability in sandboxes, and ships remediation pull requests. In benchmark testing, the system achieved 72% recall on 50 real-world vulnerabilities at 30% lower cost than the next most accurate alternative.
CognitionJun 30Cognition introduces Devin Fusion, a hybrid-model harness for agentic coding that reduces the cost of Fable-level intelligence by 35%. The system runs a cost-effective sidekick agent in parallel with a frontier model, using dynamic routing to shift tasks between them based on complexity. Devin Fusion is available now within the Devin platform.
CognitionJun 24Cognition added a testing workflow to Devin that automates end-to-end verification. After creating a pull request, Devin generates a test plan for human approval. Once approved, the agent executes the tests, records a screen-captured video with a visual QA checklist, and attaches the recording to the session for immediate verification without requiring manual branch pulls.
CognitionJun 18Cognition integrated security review into every Devin Review pull request. The agent automatically scans code for vulnerabilities like authentication bypasses and business-logic flaws that pattern-matching tools miss. It ranks findings by severity, tags them with CWE IDs, explains the exploit path, and drafts a merge-ready pull request to remediate the issue without requiring manual intervention.
Frequently asked questions