--- title: Schema harness reaches ~99% on ARC-AGI-3 public benchmark description: A new harness called Schema enables frontier LLMs to score ~99% on the ARC-AGI-3 public benchmark (25 games), compared to a 42.83% baseline for Claude Code... canonical: https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Schema harness reaches ~99% on ARC-AGI-3 public benchmark | daily.dev og:description: A new harness called Schema enables frontier LLMs to score ~99% on the ARC-AGI-3 public benchmark (25 games), compared to a 42.83% baseline for Claude Code... og:url: https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc og:image: https://api.daily.dev/og/posts/JztbNgJRC.png og:image:alt: Schema harness reaches ~99% on ARC-AGI-3 public benchmark 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. # Schema harness reaches ~99% on ARC-AGI-3 public benchmark **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary A new harness called Schema enables frontier LLMs to score ~99% on the ARC-AGI-3 public benchmark (25 games), compared to a 42.83% baseline for Claude Code without it. The approach has models write each game's mechanics as an executable, verifiable program rather than guessing outputs directly — jointly resolving state grounding and mechanism discovery in a single editable program. Search-based planning then runs inside that simulation. Results suggest a division of labor: the harness reduces the cost of using a discovered theory, while the underlying model determines how quickly a clean theory is found. Caveats include self-reported numbers, no independent ARC Prize verification yet, and the benchmark being the public 25-game set rather than a held-out evaluation. ## Content A new agent harness called Schema has pushed frontier LLMs to roughly 99% on the ARC-AGI-3 public benchmark — up from a Claude Code baseline of 42.83%. The core idea is straightforward: instead of having a model guess at each puzzle, Schema makes it act like a physicist. For every game, the model writes the mechanics as an executable program, verifies that program against recorded game history, then uses search to plan inside it. The result is a system that jointly solves two problems that usually get treated separately — figuring out what objects and variables exist (state grounding) and figuring out how state changes over time (mechanism discovery). The numbers are striking. On 25 public ARC-AGI-3 games, Schema with Claude Opus 4.8 and Fable 5 hits 98.98% RHAE. With GPT-5.6 Sol it reaches 95.35%. The controlled comparison against the Claude Code baseline makes the harness's contribution hard to dismiss. One finding worth sitting with: the harness reduces the cost of *using* a theory, but the underlying model still determines the cost of *discovering* one. Better models find better programs faster; the harness just makes sure those programs get tested and exploited systematically. That's a useful separation — it means harness improvements and model improvements compound rather than substitute. A few caveats. These results are self-reported and haven't been independently verified by ARC Prize yet. The evaluation is on the 25 public games, not a held-out set. The broader point, which several researchers have been making for a while, is that harnesses are compositional generalizers. The same structure — write a model of the world, verify it, search inside it — can in principle transfer across problem types. Whether that holds outside ARC-AGI-3 is the interesting open question. ## Questions this post answers ### What score does the Schema harness achieve on the ARC-AGI-3 public benchmark? The Schema agent harness reaches roughly 99% on the ARC-AGI-3 public benchmark, up from a Claude Code baseline of 42.83%. Specifically, Schema combined with Claude Opus 4.8 and Fable 5 hits 98.98% RHAE across 25 public games, while pairing it with GPT-5.6 Sol reaches 95.35%. These results are self-reported and not yet independently verified by ARC Prize. _Track how harness-driven leaps like this reshape agent benchmarking by following AI agent news on daily.dev._ ### How does the Schema harness improve LLM performance on ARC-AGI-3 puzzles? Instead of having a model guess at each puzzle, Schema has the model write the game's mechanics as an executable program, verify that program against recorded game history, then use search to plan within it. This jointly solves state grounding (what objects and variables exist) and mechanism discovery (how state changes over time), which are usually treated as separate problems. _Developers designing agent harnesses can follow techniques like program-verified planning on daily.dev._ ### Are the Schema harness's ARC-AGI-3 benchmark results independently verified? No, the results are self-reported and have not yet been independently verified by ARC Prize. The evaluation also covers only the 25 public ARC-AGI-3 games rather than a held-out test set, so the reported near-99% scores should be treated as preliminary until confirmed. _Stay cautious about self-reported AI benchmark claims by tracking verification updates on daily.dev._ --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc) ```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":"Schema harness reaches ~99% on ARC-AGI-3 public benchmark","url":"https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc"},"datePublished":"2026-07-16T19:50:51.082Z","dateModified":"2026-09-13T19:37:00.077Z","description":"A new harness called Schema enables frontier LLMs to score ~99% on the ARC-AGI-3 public benchmark (25 games), compared to a 42.83% baseline for Claude Code...","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":0,"discussionUrl":"https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm","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":"Schema harness reaches ~99% on ARC-AGI-3 public benchmark"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/schema-harness-reaches-99-on-arc-agi-3-public-benchmark-jztbngjrc#faq","mainEntity":[{"@type":"Question","name":"What score does the Schema harness achieve on the ARC-AGI-3 public benchmark?","acceptedAnswer":{"@type":"Answer","text":"The Schema agent harness reaches roughly 99% on the ARC-AGI-3 public benchmark, up from a Claude Code baseline of 42.83%. Specifically, Schema combined with Claude Opus 4.8 and Fable 5 hits 98.98% RHAE across 25 public games, while pairing it with GPT-5.6 Sol reaches 95.35%. These results are self-reported and not yet independently verified by ARC Prize. Track how harness-driven leaps like this reshape agent benchmarking by following AI agent news on daily.dev."}},{"@type":"Question","name":"How does the Schema harness improve LLM performance on ARC-AGI-3 puzzles?","acceptedAnswer":{"@type":"Answer","text":"Instead of having a model guess at each puzzle, Schema has the model write the game's mechanics as an executable program, verify that program against recorded game history, then use search to plan within it. This jointly solves state grounding (what objects and variables exist) and mechanism discovery (how state changes over time), which are usually treated as separate problems. Developers designing agent harnesses can follow techniques like program-verified planning on daily.dev."}},{"@type":"Question","name":"Are the Schema harness's ARC-AGI-3 benchmark results independently verified?","acceptedAnswer":{"@type":"Answer","text":"No, the results are self-reported and have not yet been independently verified by ARC Prize. The evaluation also covers only the 25 public ARC-AGI-3 games rather than a held-out test set, so the reported near-99% scores should be treated as preliminary until confirmed. Stay cautious about self-reported AI benchmark claims by tracking verification updates on daily.dev."}}]} ```