--- title: FrontierCode 2.0: Cognition's coding benchmark tops out... description: Cognition released FrontierCode 2.0, a coding benchmark that judges whether a PR should be merged, not just whether it meets spec. Prompts are deliberately... canonical: https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: FrontierCode 2.0: Cognition's coding benchmark tops out at 45.5% | daily.dev og:description: Cognition released FrontierCode 2.0, a coding benchmark that judges whether a PR should be merged, not just whether it meets spec. Prompts are deliberately... og:url: https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd og:image: https://api.daily.dev/og/posts/wOazQ82QD.png og:image:alt: FrontierCode 2.0: Cognition's coding benchmark tops out at 45.5% og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # FrontierCode 2.0: Cognition's coding benchmark tops out at 45.5% **[Collections](https://daily.dev/sources/collections)** · 2 min read · 2 upvotes · 1 comments ## Summary Cognition released FrontierCode 2.0, a coding benchmark that judges whether a PR should be merged, not just whether it meets spec. Prompts are deliberately underspecified so models must use judgment instead of following instructions step by step. The top model scores 45%, or 45.5% by Joshua_WD's reading of the leaderboard, so the benchmark is far from saturated. Joshua_WD noted that higher reasoning effort led agents to change code nobody asked them to touch, which counts against a model under mergeability standards. ## Content Cognition has released FrontierCode 2.0, a coding benchmark that tests whether a model can go beyond the spec. Most coding evals check whether a model can write code that meets a given spec. FrontierCode 1.0 asked something different: would you actually merge the PR? Version 2.0 raises the bar. The question is now not only whether you would merge it, but whether you should. ## What changed The tasks are more varied, have larger scopes, and are held to higher mergeability standards. The biggest change is that prompts are underspecified on purpose. They leave out many engineering decisions, so a model has to use its own judgment instead of following instructions step by step. ## Results The benchmark is nowhere near saturated. The top model scores 45%, or 45.5% by Joshua_WD's reading of the leaderboard. Joshua_WD flagged one detail he keeps coming back to: higher reasoning effort led agents to change code nobody asked them to touch. That's a counterintuitive result. More thinking didn't make the agents more disciplined. In a benchmark that judges whether a PR should be merged, unrequested changes count against a model. I find that more interesting than the headline score. Getting a pass rate under 50% is easy to read as "models are bad at this." The scope-creep finding suggests something more specific: the agents can write code, but they struggle to tell when to stop. ## Questions this post answers ### What is Cognition's FrontierCode 2.0 benchmark and what does it measure? FrontierCode 2.0 is a coding benchmark from Cognition that tests whether an AI-generated pull request should be merged, not just whether it meets the written spec. Unlike version 1.0, which asked whether you would merge the PR, version 2.0 asks whether you should. Prompts are deliberately underspecified so models must use engineering judgment rather than follow instructions literally. _Developers evaluating coding agents for real PR work can track benchmark results like this on daily.dev._ ### What is the top score on the FrontierCode 2.0 coding benchmark? The top model scores around 45%, reported as 45.5% in one reading of the leaderboard, meaning the benchmark is far from saturated. Tasks in version 2.0 are more varied, larger in scope, and held to higher mergeability standards than in FrontierCode 1.0. _Anyone comparing coding agent capabilities can follow benchmark updates like this on daily.dev._ ### Why do AI coding agents make unrequested changes when given more reasoning effort? On the FrontierCode 2.0 benchmark, higher reasoning effort led agents to change code nobody asked them to touch, a counterintuitive result since more thinking did not make the agents more disciplined. Since the benchmark judges whether a PR should be merged, unrequested changes count against a model, suggesting agents can write code but struggle to know when to stop. _Teams deciding how much autonomy to give coding agents can weigh findings like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@idoshamun** · 0 upvotes > I really hope all these benchmarks will allow us to better evaluate models. It's so tiring to keep running around every time there's a new model and check where it fails --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#vibe-coding](https://daily.dev/tags/vibe-coding), [#code-review](https://daily.dev/tags/code-review) [View this post on daily.dev](https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"FrontierCode 2.0: Cognition's coding benchmark tops out at 45.5%","url":"https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd"},"datePublished":"2026-10-30T22:30:51.192Z","dateModified":"2026-10-30T22:31:28.579Z","description":"Cognition released FrontierCode 2.0, a coding benchmark that judges whether a PR should be merged, not just whether it meets spec. Prompts are deliberately...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"ai,ai-agents,vibe-coding,code-review","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"FrontierCode 2.0: Cognition's coding benchmark tops out at 45.5%"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd","comment":[{"@type":"Comment","text":"I really hope all these benchmarks will allow us to better evaluate models. It’s so tiring to keep running around every time there’s a new model and check where it fails","datePublished":"2026-10-31T07:22:22.456Z","url":"https://daily.dev/posts/wOazQ82QD#c-VWXhff7aN","author":{"@type":"Person","name":"Ido Shamun","url":"https://daily.dev/idoshamun","image":"https://media.daily.dev/image/upload/s---xy_OAwk--/f_auto,q_auto/v1703781380/avatars/avatar_28849d86070e4c099c877ab6837c61f0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/frontiercode-2-0-cognition-s-coding-benchmark-tops-out-at-45-5--woazq82qd#faq","mainEntity":[{"@type":"Question","name":"What is Cognition's FrontierCode 2.0 benchmark and what does it measure?","acceptedAnswer":{"@type":"Answer","text":"FrontierCode 2.0 is a coding benchmark from Cognition that tests whether an AI-generated pull request should be merged, not just whether it meets the written spec. Unlike version 1.0, which asked whether you would merge the PR, version 2.0 asks whether you should. Prompts are deliberately underspecified so models must use engineering judgment rather than follow instructions literally. Developers evaluating coding agents for real PR work can track benchmark results like this on daily.dev."}},{"@type":"Question","name":"What is the top score on the FrontierCode 2.0 coding benchmark?","acceptedAnswer":{"@type":"Answer","text":"The top model scores around 45%, reported as 45.5% in one reading of the leaderboard, meaning the benchmark is far from saturated. Tasks in version 2.0 are more varied, larger in scope, and held to higher mergeability standards than in FrontierCode 1.0. Anyone comparing coding agent capabilities can follow benchmark updates like this on daily.dev."}},{"@type":"Question","name":"Why do AI coding agents make unrequested changes when given more reasoning effort?","acceptedAnswer":{"@type":"Answer","text":"On the FrontierCode 2.0 benchmark, higher reasoning effort led agents to change code nobody asked them to touch, a counterintuitive result since more thinking did not make the agents more disciplined. Since the benchmark judges whether a PR should be merged, unrequested changes count against a model, suggesting agents can write code but struggle to know when to stop. Teams deciding how much autonomy to give coding agents can weigh findings like this on daily.dev."}}]} ```