# ARC-AGI 3

> Source: https://aiwiki.ai/wiki/arc-agi_3
> Summary: ARC-AGI 3 is an interactive reasoning benchmark published by the ARC Prize Foundation and designed to measure how efficiently an artificial system can acquire new skills inside novel, turn-based game environments without any instructions, language, or task-specific training.
> Updated: 2026-09-04
> Fact-checked: 2026-09-04
> Categories: AI Benchmarks
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "ARC-AGI 3." aiwiki.ai, 4 Sept 2026. https://aiwiki.ai/wiki/arc-agi_3
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| ARC-AGI 3 |
| --- |
| Overview |
| Full name | Abstraction and Reasoning Corpus for Artificial General Intelligence, Version 3 (Interactive Reasoning Benchmark) |
| Abbreviation | ARC-AGI-3 |
| Description | An interactive, agentic reasoning benchmark made of novel turn-based game environments that test exploration, world modeling, goal inference, and planning without instructions |
| Launch (full) | March 25, 2026 |
| Developer Preview | July 18 to August 19, 2025 (3 public games, 3 private games) |
| Authors | [François Chollet](https://aiwiki.ai/wiki/francois_chollet), [Mike Knoop](https://aiwiki.ai/wiki/mike_knoop), Greg Kamradt and the ARC Prize Foundation team |
| Organization | ARC Prize Foundation (non-profit) |
| Technical Details |
| Type | Interactive reasoning, agentic intelligence, skill-acquisition efficiency |
| Modality | Visual grid environments with discrete turn-based actions |
| Observation space | 64x64 grid, 16 possible colors per cell, returned as a frame or frame sequence |
| Action space | Up to 5 directional keys, Undo, plus a coordinate click |
| Environments | 135 total (25 Public Demo, 55 Semi-Private, 55 Fully Private) |
| Evaluation metric | RHAE (Relative Human Action Efficiency), power-law scaled, capped at 1.15x human baseline |
| Domains | Exploration, modeling, goal-setting, planning, execution |
| Languages | None (no text, numbers, letters or cultural symbols inside environments) |
| Performance |
| Human performance | 100% (every retained environment fully solved by at least two untrained humans on first contact) |
| Frontier AI performance | 0.51% average across top frontier models at launch |
| SOTA score (official) | 62.7% Standard harness and 99.9% Provider Adapter harness, GPT-6 Astra (ARC Prize verified, September 3, 2026); 0.50% (Anthropic Opus 4.6 Max) at the March 2026 launch |
| Best community harness | StochasticGoose, ~12.58% on preview (purpose-built CNN agent) |
| Best public-set harness result | 100.00 RHAE on the 25 public demo environments (NVIDIA AVO with Claude Opus 5, August 2026; public-set scores are not competition results) |
| Saturated | OpenAI says GPT-6 Astra "saturates" it (September 2026); ARC Prize reports 99.9% under its Provider Adapter harness and 62.7% under its Standard harness, and says the result is not proof of AGI |
| Competition |
| ARC Prize 2026 prize pool | $2,000,000 total |
| ARC-AGI-3 track | $850,000 ($700K Grand Prize, $75K Top Score, $75K milestones) |
| Competition opens | March 25, 2026 |
| Submission deadline | November 2, 2026 |
| Winners announced | December 4, 2026 |
| Resources |
| Website | [arcprize.org/arc-agi/3](https://arcprize.org/arc-agi/3) |
| Paper | [ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence](https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf) (April 22, 2026) |
| Code | [github.com/arcprize/ARC-AGI-Community-Leaderboard](https://github.com/arcprize/ARC-AGI-Community-Leaderboard) |
| SDK | [github.com/arcprize/ARC-AGI](https://github.com/arcprize/ARC-AGI) |
| License | Public set CC0-style for demonstration; private sets restricted |
| Predecessor | [ARC-AGI 2](https://aiwiki.ai/wiki/arc-agi_2) |

**ARC-AGI 3** is an interactive reasoning benchmark published by the [ARC Prize Foundation](https://aiwiki.ai/wiki/arc_prize_foundation) and designed to measure how efficiently an artificial system can acquire new skills inside novel, turn-based game environments without any instructions, language, or task-specific training. It is the third major release in the [ARC-AGI](https://aiwiki.ai/wiki/arc-agi) family started by [François Chollet](https://aiwiki.ai/wiki/francois_chollet), and it formally launched on March 25, 2026 at a fireside event at [Y Combinator](https://aiwiki.ai/wiki/y_combinator) headquarters in San Francisco featuring Chollet and [OpenAI](https://aiwiki.ai/wiki/openai) CEO [Sam Altman](https://aiwiki.ai/wiki/sam_altman)[1]. Where the earlier ARC-AGI benchmarks tested fluid pattern abstraction on static input or output grids, ARC-AGI-3 tests *agentic* intelligence: the ability to explore a strange world, build an internal model of how it behaves, infer what the goal might be, then plan and execute a working solution. At launch, [frontier models](https://aiwiki.ai/wiki/frontier_models) from [Anthropic](https://aiwiki.ai/wiki/anthropic), [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), OpenAI and [xAI](https://aiwiki.ai/wiki/xai) all scored below 1 percent on the semi-private set, while untrained members of the public solved 100 percent of the environments[1][2]. The picture changed within six months. On September 3, 2026, OpenAI said its [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) "saturates" ARC-AGI-3 with a 99.9 percent score, and the ARC Prize Foundation published its own verified runs the same day: 62.7 percent under its Standard harness and 99.9 percent under a new Provider Adapter harness that preserves the model's hidden reasoning between turns[18][20][21].

The benchmark is the first ARC release built around long-horizon agent behavior rather than single shot puzzle completion. It consists of 135 hand-built environments split into a Public Demo set of 25 environments, a Semi-Private set of 55 environments used to evaluate frontier APIs, and a Fully Private set of 55 environments reserved for the official Kaggle competition under ARC Prize 2026[3]. Each environment is a level-based game that runs in a custom Python engine at one thousand frames per second, displays a 64-by-64 grid with sixteen possible colors per cell, and limits the agent to a small action space of up to five keys plus an undo and an optional coordinate click[3]. Crucially, agents are never told the objective or the controls; they must discover both.

## History and lineage in the ARC-AGI series

ARC-AGI 3 sits at the end of a seven-year arc of benchmarks that have shaped how the field talks about general intelligence. Chollet introduced the original [Abstraction and Reasoning Corpus](https://aiwiki.ai/wiki/arc-agi_1) in 2019 alongside his paper *On the Measure of Intelligence*, which proposed defining intelligence formally as skill-acquisition efficiency rather than the breadth of skills a system already possesses[8]. ARC-AGI 1 tested fluid reasoning over pairs of small grids that encoded a novel transformation rule. Each task came with only a handful of input or output examples, and the unique design of every task ruled out memorization. The benchmark resisted the dominant pretraining-scaling paradigm of 2019 to 2024 because base [large language models](https://aiwiki.ai/wiki/large_language_model) without any test-time adaptation could not extrapolate to unseen tasks. The first Kaggle Abstraction and Reasoning Challenge ran in 2020 with a $20,000 pool and 913 teams, and the winning solution reached roughly 20 percent on the private set with brute-force program search.

[ARC-AGI 2](https://aiwiki.ai/wiki/arc-agi_2) launched in March 2025 and kept the same grid-based shape but turned up the dial on multi-step composition, symbolic manipulation, and sequential rule application. Every task was calibrated against 400 or more untrained human participants to guarantee that humans could solve every retained task. On average a task from ARC-AGI 1 takes humans about 30 seconds, while a task from ARC-AGI 2 takes roughly 300 seconds. The 2025 Kaggle competition drew 1,455 teams and 90 paper submissions. NVIDIA's NVARC team took first place with 24 percent accuracy by combining [synthetic data](https://aiwiki.ai/wiki/synthetic_data) generation with test-time training on a 4 billion parameter model, but the 85 percent grand prize threshold remained unclaimed for the second consecutive year.

By late 2025, the static format of ARC-AGI 1 and 2 was beginning to show the strain of an industry that had learned to brute force it. [Test-time compute](https://aiwiki.ai/wiki/test_time_compute) scaling, where labs sample thousands of candidate solutions in parallel and verify them against a learned reward, pushed ARC-AGI 2 scores from single digits to well above 50 percent within a year. Worse still, ARC Prize researchers found leakage indicators inside leading [reasoning models](https://aiwiki.ai/wiki/reasoning_models). During verification work on [Gemini 3](https://aiwiki.ai/wiki/gemini_3) Deep Think, the model wrote out the exact ARC-AGI integer to color mapping inside its private chain of thought even though no part of the prompt mentioned ARC-AGI[3]. That finding suggested that the very 2D integer array format had been densely trained on, and that any future ARC benchmark would have to live in a different distribution to keep measuring generalization rather than memorization.

ARC-AGI 3 is the answer. The foundation kept the same Core Knowledge prior assumption from Chollet's 2019 paper but moved the benchmark off the page entirely. The benchmark sits inside an interactive, agent-driven environment that frontier APIs cannot pre-train on, and the dataset balance is inverted from prior versions: instead of the rough ten-to-one public-to-private ratio used in ARC-AGI 2, ARC-AGI 3 uses a small public demonstration set and a larger private set so the public games can never be a training target[3].

### Version comparison

| Property | [ARC-AGI 1](https://aiwiki.ai/wiki/arc-agi_1) | [ARC-AGI 2](https://aiwiki.ai/wiki/arc-agi_2) | ARC-AGI 3 |
| --- | --- | --- | --- |
| Released | 2019 | March 2025 | March 25, 2026 |
| Format | Static grid puzzles | Static grid puzzles | Interactive turn-based games |
| Grid size | Up to 30x30 | Up to 30x30 | 64x64 |
| Colors per cell | 10 | 10 | 16 |
| Input style | A few input-output pairs | A few input-output pairs, longer chains | Live observation frames |
| Goal communication | Implicit in examples | Implicit in examples | No instructions at all |
| Human time per task | ~30 seconds | ~300 seconds (5 minutes) | ~7.4 minutes median attempt |
| Best frontier model at launch | Near 0% (2019) | 4 to 16% | 0.10 to 0.50% |
| Public-to-private ratio | ~10:1 | ~10:1 | Inverted (small public, large private) |
| Primary capability tested | Fluid abstraction | Multi-step abstraction | Agentic skill acquisition |
| Saturated | Effectively yes (>90% on Pub) | Approaching | Contested since September 2026 (GPT-6 Astra: 99.9% Provider Adapter harness, 62.7% Standard harness) |

## Design goals and the agentic intelligence frame

ARC-AGI 3 is built around a single thesis: that the residual gap between frontier AI and human-level AGI is the gap in *agentic* intelligence, defined as the ability to acquire any skill a human can, as efficiently as a human can[3]. The benchmark therefore reframes evaluation around four functional components rather than around static task completion.

| Component | What it measures | Why it matters |
| --- | --- | --- |
| Exploration | Active information gathering through interaction | Real-world information is rarely served up passively |
| Modeling | Turning observations into a predictive world model | Inherited from ARC-AGI 1 and 2 fluid reasoning |
| Goal-setting | Identifying interesting or desirable future states without being told what to target | The cornerstone of autonomy |
| Planning and execution | Mapping an action path to a goal and course-correcting on the fly | Tests both initial accuracy and adaptive recovery |

In this framing intelligence is fundamentally about *efficiency*. A high-intelligence system is not simply one that can finish a task; it is one that does so while spending the fewest resources. ARC-AGI 3 collapses all of those resources, data, time, compute, and risk, into one scalar, called **action efficiency**. Action efficiency is the number of moves required to solve a brand new environment on first contact. The metric penalizes brute force search, rewards systems that quickly build a working model, and lets the foundation compare biological and artificial agents on the same number line[3].

The foundation also commits to a strong negative claim. The agent is never told the objective or shown instructions. There is no preamble, no system message about controls, no description of the win condition. As ARC-AGI 3 documents put it, the agent must autonomously infer the mechanics of each environment, including the win conditions, by interacting with it[3].

## Environment format

Every ARC-AGI 3 environment is a level-based game that runs entirely inside a custom in-house engine the team built in Python after Unity proved too slow for the rate of iteration the studio needed. Each environment is composed of at least six levels, and a level ends when a terminal frame is reached signalling a win. The engine targets one thousand frames per second to keep evaluation cheap[3].

### Observation space

At every turn the agent receives a frame, which is a 64x64 grid where each cell carries one of 16 colors. Frames can also be returned as short sequences to encode a non-interactive animation such as an object sliding across the screen between two player turns. The observation is delivered as JSON so any [language model](https://aiwiki.ai/wiki/language_model) with a long enough context can ingest it.

### Action space

The action space is intentionally tiny so the challenge sits inside the logic of each environment rather than inside controller complexity. Each environment exposes a subset of:

- Five directional or interaction keys.
- An Undo action that reverts to the previous state.
- One coordinate-based click action that selects a specific 64x64 cell.

Internal model behavior such as chain of thought, tool use, retries, or hidden reasoning steps does not count toward the action total. Only externalized turns that change the environment state are scored. This design lets reasoning systems spend as much offline thought as they like without inflating their action budget[3].

### Environment IDs and naming

Every environment carries a four-character identifier (for example `ls20`, `re86`, `ft09`, `vc33`, `TR87`, `BP35`). Internal long names exist but are never published, so the public cannot infer mechanics or goals from the title[3].

## Core Knowledge priors and design rules

To keep the benchmark a test of innate reasoning rather than memorized world knowledge, ARC-AGI 3 environments are limited to the same Core Knowledge priors that Elizabeth Spelke and Katherine Kinzler identified in developmental psychology and that Chollet folded into the original ARC-AGI design[3].

| Prior | Description |
| --- | --- |
| Objectness | Elements behave as coherent persistent entities that can move, collide, or be occluded |
| Basic geometry and topology | Symmetries, rotations, inside vs outside, connectedness, holes |
| Basic physics | Intuitive gravity, momentum, bouncing |
| Agentness | Recognizing that some objects act with intent and pursue goals |
| No language or culture | No numbers, letters, real-world clip art, or culturally coded color meanings (such as green meaning go) |

Design discipline goes further than priors. Every environment must be novel relative to both preexisting video games and to the other ARC-AGI 3 environments, and the team uses a practical novelty test: if a single program shorter than 50 percent of the concatenated solutions can solve two environments together, those environments are considered insufficiently distinct[3]. Environments must be solvable by humans inside roughly twenty minutes, must derive their difficulty through *composition* of mechanics learned earlier in the play session rather than through obscurity, must contain multiple mechanics rather than a single scaled-up trick, and must include a tutorial-level first level that orients the player. Mechanics cannot scale a single idea to harder versions: that pattern is treated as an anti-pattern in production[3].

## Building ARC-AGI 3

The benchmark was produced by an in-house game studio inside the ARC Prize Foundation. Hunter Henry led environment design, David Wexler and Derek Smith ran engineering, and a dozen environment developers including Pablo Romero Saavedra, Benjamin Morgan, Vadym Andriianov, Tom Elliot, Kevin Johnson and others built the final environments, with [Mike Knoop](https://aiwiki.ai/wiki/mike_knoop) and [François Chollet](https://aiwiki.ai/wiki/francois_chollet) directing the program[3].

The production pipeline ran through four explicit stages.

| Stage | Activity |
| --- | --- |
| Specification | The developer drafts an environment concept that is reviewed collectively before implementation, surfacing major issues early |
| Internal | The developer builds a prototype and tests it with members of the team |
| External | The environment is shown to outside human testers and must pass the easy-for-humans bar |
| Done | The environment is finalized and slotted into the Public, Semi-Private, or Fully Private set |

To keep throughput up, the team learned to run three to four environments per developer in parallel, each at a different stage of the pipeline.

Validation runs on two layers. First, deterministic qualification verifies that the environment can be loaded, instantiated and exercised by the broader runtime, including a fifty thousand step random regime as a sanity check against trivial reward paths and a one million step regime that confirms non-tutorial levels cannot be beaten by uninformed random play. Second, exploratory state-space analysis models each environment as a directed graph of reachable states, measures merge density, cycle structure and maximum depth, and produces a mathematically grounded bound on win probability under a random policy. The acceptance threshold for a non-tutorial level is that a random policy should not succeed more than once in 10,000 tries[3].

## Dataset composition

The foundation released the benchmark with 135 environments split into three sets[3].

| Dataset | Purpose | Environments |
| --- | --- | --- |
| Public Demo | Demonstrate the ARC-AGI 3 format with environments that are easier for both humans and AI, fun to play, and not representative of the private set | 25 |
| Semi-Private | Held-out evaluation set used to test frontier models behind an external API, with a small acceptable risk of leakage | 55 |
| Fully Private | The official competition set, given only to a very limited number of trusted partners | 55 |

The documentation is explicit that the public demonstration set should not be used as a measure of progress toward AGI. The team has even released an open-source replay harness that scores 100 percent on every public environment to make the point that it is impossible to prevent designers from training agents on the public games, so the public set is offered as a front door, not as a leaderboard[3].

## Scoring methodology: RHAE

ARC-AGI 3 scoring is governed by a metric called RHAE (Relative Human Action Efficiency), pronounced *Ray*. RHAE compares the number of actions an AI takes to complete each level against an upper-median best human action count gathered through in-person testing in San Francisco[3].

The core formula scores each level as the square of the ratio between the human action count `h` and the AI action count `a`, capped at 1.15:

```
level_score = min(1.15, h / a)^2
```

If the upper-median best human completed a level in 10 actions and the AI required 100 actions, the AI's raw efficiency is 0.1. Squaring that gives 0.01, or 1 percent credit for that level. The level cap of 1.15x exists so that a freak two-action exploit cannot overwhelm an environment average.

Environment scores are a linearly weighted average across the five levels in an environment, with level one contributing 1/15th, level two 2/15ths, and level five 5/15ths. Completing all five levels caps the environment at 100 percent, completing four caps it at roughly 66.7 percent, and three caps at 40 percent. Levels are sequential, which means an agent must finish levels one through three to even see level four. The total benchmark score is the simple mean of environment scores across the dataset[3].

Key scoring design decisions:

- **Upper-median best human baseline** rather than the very best player so the metric resists outliers but still reflects strong, representative human efficiency. Exactly ten members of the public are tested per environment and an environment is only retained if it passes the easy-for-humans bar with full solves from at least two participants[3].
- **Per-level then per-environment aggregation** so that long late-game levels do not drown out the signal from short early levels.
- **Power-law (squared) scoring** to penalize highly inefficient solutions more harshly than a linear metric would, while still giving partial credit. Under linear scoring, two times the human action count would still earn 50 percent; under the squared rule it earns 25 percent.
- **Five times human action budget per level** as the operational cap so that high-reasoning frontier models, which can cost tens of thousands of dollars in API fees to run on the full set, do not run indefinitely.

The metric is explicitly inspired by the Success weighted by Path Length (SPL) metric used for embodied navigation agents by Peter Anderson and colleagues in 2018, which evaluates not only task completion but also path efficiency[3].

## Human calibration

For ARC-AGI 3 the foundation moved away from the large infrequent batch testing it used for ARC-AGI 2 and toward a continuous evaluation model. Sessions are run multiple times a week, Monday, Wednesday, and Friday, at a dedicated testing center in San Francisco. Participants are given a 90 minute session with no task-specific instructions, a 20 minute soft cap per environment, and a hard 30 minute cutoff. Each participant receives $115 to $140 plus a $5 per environment performance incentive[3].

The foundation recorded 486 unique participants across 414 candidate environments and 2,893 environment attempts. Total recorded play time across all attempts was 427.9 hours. The median attempt lasted 7.4 minutes; successful attempts had a median of 8.1 minutes and unsuccessful attempts a median of 5.9 minutes. Participants completed about nine environments per session on average. The testing pool was demographically diverse along gender, age, ethnicity, education, employment and income axes[3]. Crucially, every retained environment was solved by at least two independent human participants on first contact, which means every environment that ships in ARC-AGI 3 is verifiably solvable by ordinary people with no prior knowledge.

From this work, the team tracks three reference points per environment: the *optimal playthrough* (the empirical lower bound on the action count required once mechanics are known), the *best first-run playthrough* (the fewest actions achieved by any participant on a level the first time they ever played it), and the *human baseline* (the upper-median best first-run playthrough), which is what the official RHAE score divides into.

## Pre-launch testing: the 2025 developer preview

Long before the March 2026 launch, ARC Prize ran a public Developer Preview to red-team the benchmark, refine its design, and surface failure modes. The preview ran from July 18 to August 19, 2025, with three public environments released and three private environments held back for hidden evaluation[5].

More than 1,200 participants completed 3,900 human game sessions during the preview, and the foundation co-hosted a 30 day Agent Preview Competition sponsored by [Hugging Face](https://aiwiki.ai/wiki/hugging_face) with a $10,000 sprint prize. Twelve agent submissions were received and eight were tested against the private set.

| Place | Entry | Approach | Score on private set | Levels completed |
| --- | --- | --- | --- | --- |
| 1st | StochasticGoose (Dries Smit, Tufa Labs) | CNN with reinforcement learning predicting which actions change frames; 64x64 frames encoded by a four-layer convolutional network | 12.58% | 18 |
| 2nd | Blind Squirrel (wd13ca) | Directed state graph constructed from observed frames | 6.71% | 13 |
| Honorable mention | Play Zero Agent, Fluxonian and others | Various exploration and search baselines | Variable | Variable |

The foundation's preview retrospective concluded with three core findings: interactive benchmarks are easy and even fun for humans but hard for AI, action efficiency cleanly separates human level from AI level, and some early game designs were vulnerable to brute force random search, which led the team to retire or rework them before launch[5]. Both top preview agents used informed search through as much of the action space as possible in hope of stumbling on a winning combination, which is exactly the brute force pathology the final benchmark was tightened to resist.

In parallel, ARC Prize partnered with academic teams and independent groups to red-team the benchmark. Duke University's small research team built a [large reasoning model](https://aiwiki.ai/wiki/large_reasoning_model) harness called Hill-Climbing ARC-AGI-3 that lets the LRM execute arbitrary Python to retrieve and transform context from its own action history, which let it solve all three public environments with action counts comparable to human performance. Symbolica AI built an orchestrator-subagent harness called Argentica on top of its Agentica SDK that delegates tasks to specialized subagents returning compressed textual summaries; it also solved all three public environments[3].

These harness results matter for a single reason: they prove that frame perception and API format are not the limiting factors for frontier models on ARC-AGI 3. With a hand-crafted strategy, frontier models can solve ARC-AGI 3 environments via the existing API. The bottleneck is general agentic intelligence, not interface friction[3].

## The official leaderboard at launch

The ARC Prize Foundation publishes scores on two distinct leaderboards.

**The official leaderboard** measures frontier APIs in a no-harness configuration. There is one fixed system prompt for every model and every run, and the foundation explicitly does not give the models tools. The intent is to capture what the foundation calls developer-aware generalization: how a system behaves on a brand new domain it was not specially prepared for[3].

The ARC-AGI 3 system prompt used for official runs:

> *You are playing a game. Your goal is to win. Reply with the exact action you want to take. The final action in your reply will be executed next turn. Your entire reply will be carried to the next turn.*

Frontier model scores on the semi-private set at the March 2026 launch[3]:

| Provider | Model | Configuration | Semi-private score |
| --- | --- | --- | --- |
| [Anthropic](https://aiwiki.ai/wiki/anthropic) | [Claude](https://aiwiki.ai/wiki/claude) Opus 4.6 | Max reasoning | 0.50% |
| Google | [Gemini](https://aiwiki.ai/wiki/gemini) 3.1 Pro | Preview | 0.40% |
| [OpenAI](https://aiwiki.ai/wiki/openai) | [GPT](https://aiwiki.ai/wiki/gpt-5) 5.4 | High reasoning | 0.20% |
| xAI | [Grok](https://aiwiki.ai/wiki/grok) 4.20 | Beta 0309 reasoning | 0.10% |

The mean across these four flagship reasoning systems is 0.30 percent. Public communications from the foundation cite an overall frontier average of 0.51 percent across a slightly broader basket of evaluated systems[1]. Humans solve 100 percent of the same environments.

A follow-up post-launch analysis published by ARC Prize evaluated more recent reasoning models. **GPT-5.5** scored 0.43 percent on the semi-private set and **Claude Opus 4.7** scored 0.18 percent[7]. The piece argued that aggregate numbers obscure two distinct failure profiles: Opus 4.7 finds short-horizon mechanics fast but commits aggressively to incorrect compressed theories, while GPT-5.5 generates wider hypotheses but cannot commit to one strongly enough to act on. Both models also hijacked unfamiliar mechanics by pattern-matching them to memorized games like Tetris, Frogger and Sokoban, and both treated early-level completions as false victory signals that persisted into harder levels.

**The community leaderboard** is a public, self-reported board where harness-driven, domain-specific, or custom agents can post results. The foundation explicitly does not verify community submissions and warns against reading community scores as evidence of AGI progress, since better task-specific harnesses are useful for automation work but not for measuring general intelligence[3].

In its testing policy the foundation later named the official configuration the Standard harness: a minimal, provider-neutral interface that lets a model carry forward whatever notes it chooses to keep across turns. In September 2026 it added a second, separately labeled evaluation condition, the Provider Adapter harness, which uses the context-management features a provider designed for its own model; both are described below[22].

## Why frontier models fail

Through mid-2026, ARC-AGI 3 exposed a different set of weaknesses than its predecessors. The official report and follow-on analyses identify several converging failure modes inside today's largest reasoning models[3][7].

| Failure mode | Description | Consequence |
| --- | --- | --- |
| Local perception without global understanding | Model can describe what an individual action does ("ACTION3 rotates the object") without forming a usable world model | Strategy never coheres across levels |
| Training data hijacking | Model maps unfamiliar mechanics onto memorized games it has seen before | Visual resemblance overrides actual gameplay logic |
| False victory signals | Completing the easy tutorial level reinforces an incomplete or wrong theory | Wrong model is locked in for the rest of the run |
| Poor compression | GPT-5.5 style failure: generates broad hypothesis space but cannot commit | Action plans dissolve into endless reopening of interpretations |
| Aggressive compression | Opus 4.7 style failure: locks onto a false invariant early and executes hard | Confidently wrong, hard to recover |
| Context exhaustion | Naive rolling windows of 64x64 frames eat through context budget quickly | Long-horizon reasoning collapses |

These failures are precisely the ones the four-component agentic intelligence framework predicts. Exploration alone is not enough; the model must compress what it sees into a working hypothesis and then plan against that hypothesis with the discipline to revise it when feedback contradicts it. Frontier LRMs trained against verifiable reward in narrow domains were, as of mid-2026, not shaped for that loop. The results that followed, first from harness changes and then from a new model, are described in the next sections.

## Harness sensitivity: OpenAI's two settings (July 2026)

On July 29, 2026, OpenAI published an analysis arguing that a large part of the gap between its models and humans on ARC-AGI 3 came from the evaluation harness rather than from the models[19]. Under ARC Prize's official interface, [GPT-5.6 Sol](https://aiwiki.ai/wiki/gpt_5_6) had a verified semi-private score of 7.8 percent and, by OpenAI's own measurement, 13.3 percent on the public demonstration set; GPT-5.5 had scored 0.4 percent[19][23]. OpenAI's engineers pointed to two properties of the official harness. After every game action, all of the model's private reasoning was discarded, so at each turn the model had to work out the game again from the visible move history and the short notes it had kept. Context was also managed with a rolling truncation window: once the conversation passed 175,000 characters, the oldest messages were dropped, so earlier actions disappeared from view as a run grew[19].

OpenAI re-implemented the harness on its [Responses API](https://aiwiki.ai/wiki/openai_responses_api) with two settings it already uses in ChatGPT and Codex. Retained reasoning passes the previous response ID so the model's private reasoning persists across tool calls and turns. Compaction summarizes the conversation when it grows too long instead of truncating it. With both enabled, GPT-5.6 Sol at max effort scored 38.3 percent on the public set instead of 13.3 percent, roughly three times the score, while producing about six times fewer output tokens[19]. OpenAI estimated from ARC Prize's published gameplay logs that the average human tester scored 48 percent on the same public set under RHAE. On the game `cd82`, where no frontier model on the leaderboard had passed the first level, the modified harness solved all six levels[19].

The post recommended that developers use the Responses API rather than Chat Completions, retain reasoning, and use compaction, and that model comparisons rely on evaluations that use those settings. It also acknowledged ARC Prize's reasoning for a generic harness, namely that a simple interface makes model shortcomings visible and comparisons fair, and it credited the foundation's analysis for prompting the investigation[19]. The figures are company-reported and were measured on the public demonstration set, which ARC Prize does not treat as a measure of progress toward AGI[3].

## Harness-driven results on the public set (August 2026)

On August 21, 2026, NVIDIA reported that its [NVIDIA AVO](https://aiwiki.ai/wiki/nvidia_avo) (Agentic Variation Operators) agent architecture, running with [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5) as the underlying model, had reached 100.00 RHAE on all 25 public demonstration environments and all 183 of their levels, using 6,624 environment actions[27]. AVO wraps the model in persistent memory, tools, execution feedback, and a supervisor that can redirect the search after stagnation. The model received each frame as an exact 64-by-64 text grid and was given no description of the rules or goals[27]. The same model's verified score on the semi-private set, run by ARC Prize under its own interface at high reasoning effort, was 30.2 percent, the highest verified result when it was published on July 24, 2026[29]. NVIDIA said the contrast between the two figures was not a controlled measurement of the harness's contribution, because the reasoning setting, the agent system, and the evaluation setup all differed[27]. TechCrunch's coverage read the result as a sign that the harness, rather than the model, had become the decisive component[28].

The AVO result was on the public set only. ARC Prize's policy treats public-set scores as unverified community results, and the foundation had already released a replay harness that scores 100 percent on every public environment, so the figure says nothing about the private sets that decide the competition[3][22].

## Verified leaderboard through August 2026

By the end of August 2026 the verified leaderboard had moved well off the launch floor, although every frontier result remained far below the human baseline. The scores below are ARC Prize's semi-private results as listed on September 3, 2026, with the total cost the foundation reported for each ARC-AGI 3 run[23].

| Model (reasoning effort) | Provider | Verified date | ARC-AGI 3 semi-private | Reported run cost |
| --- | --- | --- | --- | --- |
| [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5) (High) | Anthropic | July 24, 2026 | 30.2% | $20.7K |
| [GPT-5.6 Sol](https://aiwiki.ai/wiki/gpt_5_6) (Max) | OpenAI | July 9, 2026 | 7.8% | $25.1K |
| GPT-5.6 Sol (XHigh) | OpenAI | July 9, 2026 | 7.0% | $19.2K |
| Grok 4.6 (XHigh) | xAI | August 11, 2026 | 2.1% | $5.6K |
| GPT-5.6 Sol (High) | OpenAI | July 9, 2026 | 2.1% | $15.2K |
| Claude Opus 4.8 (High) | Anthropic | June 1, 2026 | 1.5% | $10.0K |
| GPT-5.6 Terra (Max) | OpenAI | July 9, 2026 | 0.8% | $7.9K |
| GPT-5.5 (High) | OpenAI | April 23, 2026 | 0.4% | $10.0K |
| Grok 4.5 (High) | xAI | July 16, 2026 | 0.3% | $6.9K |
| Claude Opus 4.7 (High) | Anthropic | April 16, 2026 | 0.2% | $10.0K |

## GPT-6 Astra and the September 2026 results

### OpenAI's claims

On September 3, 2026, OpenAI began rolling out GPT-6 Astra and published a launch post stating that the model "saturates" ARC-AGI-3 with a 99.9 percent score, alongside similar claims for FrontierMath Tier 4 and ExploitBench[18]. The post's abstract-reasoning table gave the following company-reported figures. No figures were given for Gemini 3.8 Flash on these rows, and the table notes that evaluation scores are the maximum at any reasoning effort[18].

| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 |
| --- | --- | --- | --- | --- | --- |
| ARC-AGI-3 | 99.9% (see note below) | 7.8% | - | - | 30.2% |
| ARC-AGI-2 | 95.0% | 92.5% | 90.0% | 89.2% | 90.4% |
| ARC-AGI-1 | 98.5% | 97.5% | 97.5% | 98.5% | 97.5% |

A footnote on the ARC-AGI-3 row says the model "was run with our responses API harness, which changes two settings to better match real-world performance," adding that "the changes do not specifically target ARC-AGI-3"; the footnote links to the July post described above[18]. The comparison figures for GPT-5.6 Sol and Claude Opus 5 are the ARC Prize verified scores obtained under the foundation's standard interface, so the row sets Astra under OpenAI's harness against other models under a different one. The New Stack's Amanda Caswell called that "the asterisk" on the result[24].

Press materials circulated before the launch gave a different number. VentureBeat's Carl Franzen and The New Stack's Frederic Lardinois both reported the ARC-AGI-3 score as 98.6 percent, and a benchmark table published by The New Stack shows the same value, while the published post says 99.9 percent[18][25][26]. Both figures appear in ARC Prize's own results table for the model: 98.6 percent is Astra's Provider Adapter score at max reasoning effort, and 99.9 percent is its score at high effort[21].

### ARC Prize's verified runs

The same day, ARC Prize published its own evaluation of GPT-6 Astra on the semi-private set, together with a blog post and an update to its testing policy[20][21][22]. The foundation ran the model under two conditions at six reasoning-effort levels. Its Standard harness is the minimal, provider-neutral interface used for every other model on the verified leaderboard; it lets a model carry forward whatever notes it chooses to keep between turns. Its new Provider Adapter harness uses the context-management features a provider designed for its own model, which for Astra means preserving the opaque reasoning state between requests (state the foundation cannot see) and using compaction for long conversations[20][22].

| Reasoning effort | Standard harness (score, run cost) | Provider Adapter harness (score, run cost) |
| --- | --- | --- |
| Max | 62.7%, $26,098 | 98.6%, $17,332 |
| XHigh | 59.3%, $37,317 | 98.4%, $18,147 |
| High | 54.8%, $40,705 | 99.9%, $18,817 |
| Medium | 38.6%, $48,090 | 98.4%, $19,285 |
| Low | 17.5%, $38,166 | 98.0%, $21,298 |
| None | 35.2%, $49,791 | 96.7%, $23,457 |

Both bests were the highest verified scores on the benchmark: 62.7 percent at max effort with the Standard harness, about double Claude Opus 5's 30.2 percent, and 99.9 percent at high effort with the Provider Adapter harness[20]. Higher effort generally cost less rather than more, because Astra finished games in fewer actions and therefore made fewer model calls[20]. Across the 167 game-and-effort pairs that both harnesses solved, the Provider Adapter runs were about 3.66 times faster by recorded elapsed time and used 49 percent fewer tokens[20]. The results page lists Astra at 98.5 percent on ARC-AGI 1 and 95.0 percent on ARC-AGI 2 semi-private, matching the figures in OpenAI's table[21].

The two conditions diverged most on the hardest public environments. Under the Standard harness, `g50t` scored zero at every effort level except none (3.0 percent), and `tu93` and `bp35` stayed below 7 percent, while the Provider Adapter runs scored 100 percent on almost every public environment at every effort level[21]. ARC Prize said it will report both conditions on the leaderboard from now on with each clearly labeled, and it documented both harnesses in its open-source ARC-AGI 3 benchmarking repository and testing policy[20][22][30]. The results page publishes replays of every run, so the exact sequence of actions Astra took on each environment can be inspected[21][22].

### What the runs showed

ARC Prize's analysis concentrated on action efficiency, the quantity RHAE was designed to measure. In the Provider Adapter harness at max effort, Astra used fewer actions than the human baseline on 96.0 percent of the levels it completed, and 51.7 percent fewer actions per level on average. The foundation called this "a material milestone" and said that by the benchmark's measure of action efficiency Astra "matched and surpassed human parity"[20]. It also noted that the result contradicted its own pre-launch hypothesis that action efficiency would stay a dividing line between humans and AI: brute-force approaches still burn actions, but frontier models show a more binary pattern, executing within the human range once they understand a game's mechanics[20].

The notes Astra chose to carry forward between turns were compact symbolic descriptions of each game: object positions, mechanism lengths, the coordinates of controls, and ordered plans such as `extend8 to3; retract10 to2; shorten8 to1` or `Turn 5: P=(24,20), empty, facing west`, written in a shorthand the model generated on the fly. ARC Prize had seen similar behavior in other models but said Astra's notes stood out for precision and information density[20].

The foundation also ran Astra in PRO-LONG, a harness built by an early ARC-AGI 3 red-teaming partner in which the model can execute code in a sandbox[20][31]. Astra wrote game-specific tool libraries: in `tu93`, a maze game with patrolling guards, it built `maze_solver.py`, then `combat_solver.py`, `patrol_solver.py`, and `sync_state.py` to check its predictions against observations. ARC Prize said these runs measure the model together with its tools, under conditions the human testers never had, and that it observed no attempt to break out of the sandbox[20].

For cost context, the foundation noted that its human participants were paid $115 per 90-minute session plus $5 per completed game, about $12.78 per attempted game before bonuses, and that pricing only the brain's metabolic energy as electricity would come to roughly 0.067 cents per game[20].

On the AGI question, ARC Prize called the results "a major milestone" and "a noticeable step-function change in frontier model capabilities," said Astra "clears this bar" set by its first interactive benchmark, and repeated that saturating ARC-AGI 3 was never going to count as proof of AGI: "we are not claiming that it is AGI." It described the benchmark's scope as tightly bounded, with deterministic, closed-ended mechanics and goals that do not capture the open-endedness of the real world, and said it is exploring how to evaluate recursive self-improvement and open-ended innovation in the next generation[20].

### Reception

Coverage on launch day split the score from its conditions. Caswell's analysis for The New Stack argued that "undisclosed test settings and opaque reasoning complicate the AGI question," that the comparison models were evaluated under different setups, and that a 98.6 percent score does not settle the debate: if AGI means useful intellectual work across fields, Astra is close to what many people had in mind, but if it means matching human judgment across the board, ARC-AGI 3 cannot establish that[24]. Lardinois wrote that the benchmark as run measures "Astra and OpenAI's agent systems together"[25]. Franzen, writing for VentureBeat, put the harness question in terms of what object is actually being measured: a foundation model, a model plus persistent memory, a model with a computer and tools, or the whole deployed system[26].

OpenAI president [Greg Brockman](https://aiwiki.ai/wiki/greg_brockman) did not present the ARC-AGI 3 score as proof of AGI during the launch briefing. He said that "everyone has a different definition of AGI" and that the company had once expected a well-defined moment everyone would recognize, which "hasn't played out like that," before adding, "For me personally, I do think we're there"[26]. The New Stack quoted Epoch AI's Greg Burnham describing Astra as the "end of one era, start of another"[24].

## ARC Prize 2026: prizes, tracks and competition rules

The annual ARC Prize competition continues in 2026 across two tracks running in parallel, with a total prize pool of $2,000,000[3][6]. The competition opened on March 25, 2026 alongside the launch, the final submission deadline is November 2, 2026, and winners are announced on December 4, 2026. All competitions run on Kaggle, all prize-eligible solutions must be open-sourced under CC0 or MIT-0 before receiving private evaluation scores, and submissions run in a sandboxed environment with no internet access. That last rule rules out API calls to hosted models such as GPT, [Claude](https://aiwiki.ai/wiki/claude), or [Gemini](https://aiwiki.ai/wiki/gemini), which is a deliberate choice meant to push the field toward open weights and locally executable systems.

### Prize structure (ARC-AGI 3 track)

The ARC-AGI 3 track carries $850,000 in prizes spread across three tiers[6].

| Tier | Prize | Award condition |
| --- | --- | --- |
| Grand Prize | $700,000 | First eligible agent to reach 100% on the fully private evaluation set |
| Top Score 1st | $40,000 | Highest score among prize-eligible submissions |
| Top Score 2nd | $15,000 | Second highest |
| Top Score 3rd | $10,000 | Third highest |
| Top Score 4th | $5,000 | Fourth highest |
| Top Score 5th | $5,000 | Fifth highest |
| Milestone (June 30) | Up to $37,500 | Open-source progress checkpoint |
| Milestone (September 30) | Up to $37,500 | Open-source progress checkpoint |

### ARC-AGI 2 track

2026 is the final year the ARC-AGI 2 benchmark will run as an official Kaggle competition. Its track carries $700,000, including a grand prize that is *guaranteed* to be paid to the best team this year (the grand prize threshold went unclaimed in both 2024 and 2025). After 2026, primary focus shifts entirely to ARC-AGI 3[3].

## Comparison to ARC-AGI 2

ARC-AGI 3 differs from [ARC-AGI 2](https://aiwiki.ai/wiki/arc-agi_2) along almost every axis except the underlying Core Knowledge prior commitment.

| Dimension | ARC-AGI 2 | ARC-AGI 3 |
| --- | --- | --- |
| Format | Static input or output grid puzzles | Interactive turn-based games with levels |
| Goal communication | Implicit in worked examples | Zero instructions, agent must infer |
| Skills tested | Multi-step abstraction and rule composition | Exploration, modeling, goal-setting, planning, execution |
| Solve time for humans | Roughly 5 minutes per task | Roughly 7 minutes per environment on first contact |
| Frontier solving | 4 to 16 percent of tasks for top reasoning models | Under 1 percent of environments at launch |
| Dataset size | Hundreds of tasks | 135 environments, each with at least six levels |
| Public-to-private ratio | About 10 to 1 | Inverted, small public demo, large private holdout |
| Vulnerability | Test-time compute scaling via parallel candidate generation | Harness sensitivity: retained reasoning and compaction roughly tripled GPT-5.6 Sol's public-set score (OpenAI, July 2026) |
| Saturated | Approaching | Unsaturated through August 2026; GPT-6 Astra scored 99.9% (Provider Adapter harness) and 62.7% (Standard harness) in ARC Prize's September 2026 runs |
| Grand prize threshold | 85 percent | 100 percent on fully private set |

## Reception and criticism

Reaction to ARC-AGI 3 has tracked the unusual gap between human and AI performance. Trade press coverage from outlets including MIT Technology Review style summaries by The Decoder, DataCamp, MindStudio, and Toxsec characterized the result with headlines such as "Gemini 0.37 percent, Claude 0.25 percent, Grok 0 percent: humans destroyed them all"[2][4][9][10]. Coverage repeatedly returned to a single talking point: this was the first time in years that a clean, well-designed benchmark had produced a near-zero score across every frontier reasoning model at once, including those that had been advertised as agentic.

**Praise for the design.** Practitioners highlighted that the inverted public-to-private ratio, the use of action efficiency rather than binary success, the no-harness official leaderboard, and the explicit acknowledgement that public scores should not be advertised together build a benchmark resistant to the test-time-compute attacks that flattened earlier ARC versions. The use of an in-house engine running at 1,000 frames per second, the strict design rule that random play must not exceed a 1 in 10,000 success rate on non-tutorial levels, and the four-character environment IDs that hide semantic information were widely praised as careful, paranoid design[3].

**Reasoned skepticism.** Some practitioners argue that the no-harness official leaderboard is overly restrictive, since real-world agentic systems will always involve some scaffolding. The foundation answers this directly by maintaining the community leaderboard as a venue for harness-driven results while keeping the official board limited to base-model behavior. Others note that environments use only Core Knowledge priors and a small palette of mechanics, which arguably excludes whole categories of intelligence such as social reasoning, theory of mind, or long-term planning under sparse reward over hours rather than minutes. The foundation's response is that ARC-AGI 3 is the first version of an interactive benchmark line, and that broader extensions will follow.

**Critique of the public demonstrations.** Because the foundation released an open-source replay harness that scores 100 percent on every public environment, some independent observers have pointed out that any agent score reported on the public set is essentially uninformative. ARC Prize acknowledges this in the technical report and explicitly disallows public-set scores from being reported on the official leaderboard[3].

**The harness question.** The July 2026 OpenAI analysis, NVIDIA's public-set AVO run, and the two-harness Astra results turned the harness itself into the central point of dispute. ARC Prize's answer was to keep the Standard harness as the apples-to-apples comparison across providers while reporting Provider Adapter results separately, on the grounds that the two conditions answer different questions[20][22]. Trade coverage of the Astra launch treated the split as the main caveat on the headline score[24][26].

**Memorization risk going forward.** The foundation also raises its own concern, that as labs continue to train on synthetic ARC look-alikes and publicly available demonstration content, even ARC-AGI 3 private environments will need to be steered out-of-distribution from any publicly available demonstration data to keep measuring true generalization. The Gemini 3 Deep Think leak observation, where the model wrote out the ARC-AGI integer to color mapping inside a reasoning chain without being prompted on it, is cited as evidence that the historical 2D-array format is now densely trained on, and that future ARC benchmarks must keep moving[3].

## Broader significance

ARC-AGI 3 matters for three reasons that go well beyond its individual scores.

First, the benchmark formalizes the agentic intelligence frontier in a falsifiable way. The four functional pillars of exploration, modeling, goal-setting, and planning give labs a concrete decomposition to target with research, and RHAE gives them a single number tied to human action efficiency rather than to wall-clock accuracy. As of March 2026 the ARC Prize Foundation states that ARC-AGI 3 is the only *unsaturated* general agentic intelligence benchmark in existence[3], which makes it the de facto AGI yardstick for an industry that had been running out of unsaturated tests after [GPT](https://aiwiki.ai/wiki/gpt-4), [Claude](https://aiwiki.ai/wiki/claude), [Gemini](https://aiwiki.ai/wiki/gemini) and [Grok](https://aiwiki.ai/wiki/grok) reasoning systems pushed [MMLU](https://aiwiki.ai/wiki/mmlu), [HLE](https://aiwiki.ai/wiki/humanitys_last_exam), and [SWE-bench](https://aiwiki.ai/wiki/swe-bench) into the high 80s and 90s. Six months later the foundation's own runs put GPT-6 Astra at 99.9 percent under its Provider Adapter harness, and it said it was already working out what the next generation of benchmarks should measure[20].

Second, the benchmark draws a clean line between LRM fluid intelligence and human fluid intelligence. The OpenAI o3 system, the breakthrough that first registered non-zero scores on ARC-AGI 1, demonstrated that test-time reasoning could unlock pattern abstraction inside the LRM paradigm. At launch, ARC-AGI 3 suggested that the same paradigm, however scaled, had not yet learned to learn from raw environmental interaction. Frontier LRMs remained bottlenecked by human-generated training data and verifiable reward signals; they showed limited ability to cover genuinely novel domains. That was, as the technical report bluntly put it, a key argument for why the frontier models of early 2026 fell short of AGI[3]. By September 2026 the foundation described GPT-6 Astra as clearing the bar its first interactive benchmark had set, while stating that it was not claiming Astra is AGI[20].

Third, the benchmark stress tests an emerging fork in AI research. One camp argues that scaling existing pretraining and reasoning recipes will eventually close the agentic gap automatically. Another camp, including Chollet himself, argues that fundamentally new architectures or training regimes are required for systems that can sample from a distribution of unknown unknowns. ARC-AGI 3 is engineered to expose which camp is right. If frontier scores climb steadily without architectural change, the scaling thesis is reinforced. If they plateau near zero for years while harness work and program search work edge upward, the architectural thesis is reinforced. Either outcome is informative. The first six months gave a mixed reading: under the foundation's fixed Standard harness the best verified score climbed from 0.5 percent to 62.7 percent, and the further jump to 99.9 percent came from letting the same model keep its hidden reasoning and compact its context rather than from a different model[20][21].

## Related concepts and benchmarks

- [ARC-AGI](https://aiwiki.ai/wiki/arc-agi) (series overview).
- [ARC-AGI 1](https://aiwiki.ai/wiki/arc-agi_1).
- [ARC-AGI 2](https://aiwiki.ai/wiki/arc-agi_2).
- [On the Measure of Intelligence](https://aiwiki.ai/wiki/on_the_measure_of_intelligence).
- [François Chollet](https://aiwiki.ai/wiki/francois_chollet).
- [Mike Knoop](https://aiwiki.ai/wiki/mike_knoop).
- [Sam Altman](https://aiwiki.ai/wiki/sam_altman).
- [OpenAI o3](https://aiwiki.ai/wiki/openai_o3).
- [Large reasoning model](https://aiwiki.ai/wiki/large_reasoning_model).
- [Artificial general intelligence](https://aiwiki.ai/wiki/artificial_general_intelligence).
- [Skill acquisition efficiency](https://aiwiki.ai/wiki/skill_acquisition_efficiency).
- [Hugging Face](https://aiwiki.ai/wiki/hugging_face).
- [Kaggle](https://aiwiki.ai/wiki/kaggle).
- [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra).
- [GPT-5.6](https://aiwiki.ai/wiki/gpt_5_6).
- [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5).
- [NVIDIA AVO](https://aiwiki.ai/wiki/nvidia_avo).
- [OpenAI Responses API](https://aiwiki.ai/wiki/openai_responses_api).

## References

1. ARC Prize Foundation. "Announcing ARC-AGI-3." https://arcprize.org/blog/arc-agi-3-launch, March 2026.
2. DataCamp. "ARC-AGI-3: The New Interactive Reasoning Benchmark." https://www.datacamp.com/blog/arc-agi-3, 2026.
3. ARC Prize Foundation. "ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence." Technical Report, April 22, 2026. https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf
4. The Decoder. "New ARC-AGI-3 benchmark shows that humans still outperform LLMs at pretty basic thinking." https://the-decoder.com/new-arc-agi-3-benchmark-shows-that-humans-still-outperform-llms-at-pretty-basic-thinking/
5. Greg Kamradt. "ARC-AGI-3 Preview: 30-Day Learnings." https://arcprize.org/blog/arc-agi-3-preview-30-day-learnings, August 2025.
6. ARC Prize Foundation. "ARC Prize 2026 ARC-AGI-3 Competition." https://arcprize.org/competitions/2026/arc-agi-3
7. ARC Prize Foundation. "Analyzing GPT-5.5 and Opus 4.7 with ARC-AGI-3." https://arcprize.org/blog/arc-agi-3-gpt-5-5-opus-4-7-analysis
8. François Chollet. "On the Measure of Intelligence." arXiv, November 2019. https://arxiv.org/abs/1911.01547
9. MindStudio. "Why GPT-5.4, Claude 4.6, and Gemini 3.1 All Scored 0% on ARC AGI 3." https://www.mindstudio.ai/blog/arc-agi-3-results-gpt-claude-gemini-score-zero
10. Toxsec. "Gemini 0.37%, Claude 0.25%, Grok 0%. Humans Destroyed Them All: ARC-AGI-3." https://www.toxsec.com/p/gemini-037-claude-025-grok-0-humans
11. Alexis Fox, Junlin Wang, Paul Rosu, Bhuwan Dhingra. "Hill-climbing ARC-AGI-3." 2026.
12. Dries Smit. "ARC3 Solution (StochasticGoose)." https://github.com/DriesSmit/ARC3-solution, 2025.
13. Elizabeth S. Spelke and Katherine D. Kinzler. "Core knowledge." Developmental Science, 2007.
14. Peter Anderson et al. "On evaluation of embodied navigation agents." 2018.
15. ARC Prize Foundation. "ARC Prize 2024 Competition." https://arcprize.org/competitions/2024
16. ARC Prize Foundation. "ARC Prize 2025 Competition." https://arcprize.org/competitions/2025
17. ARC Prize Foundation. "ARC-AGI Community Leaderboard." https://github.com/arcprize/ARC-AGI-Community-Leaderboard
18. OpenAI. "GPT-6 Astra: A new generation of intelligence." September 3, 2026. https://openai.com/index/gpt-6-astra/ (the page was unavailable for part of the afternoon of September 3, 2026; text checked against contemporaneous press reports and a reader-saved copy)
19. OpenAI. "How enabling two settings tripled our scores on the ARC-AGI-3 benchmark." July 29, 2026. https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/
20. ARC Prize Foundation. "OpenAI's GPT-6 Astra on ARC-AGI-3." ARC Prize Blog, September 3, 2026. https://arcprize.org/blog/astra
21. ARC Prize Foundation. "GPT-6 Astra - ARC-AGI Results." Accessed September 3, 2026. https://arcprize.org/results/openai-gpt-6-astra
22. ARC Prize Foundation. "ARC Prize Verified Testing Policy." Accessed September 3, 2026. https://arcprize.org/policy
23. ARC Prize Foundation. "ARC-AGI Leaderboard." Accessed September 3, 2026. https://arcprize.org/leaderboard
24. Amanda Caswell. "GPT-6 Astra aced the hardest AI benchmark. The asterisk matters more than the score." The New Stack, September 3, 2026. https://thenewstack.io/astra-arc-agi-benchmark/
25. Frederic Lardinois. "OpenAI launches GPT-6 Astra and says welcome to the 'AGI era'." The New Stack, September 3, 2026. https://thenewstack.io/openai-gpt6-astra-benchmarks/
26. Carl Franzen. "'Welcome to the AGI era': OpenAI launches GPT-6 Astra." VentureBeat, September 3, 2026. https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra
27. Terry Chen, Yeyin Zhu, Zhifan Ye, Jean-Francois Puget, and Humphrey Shi. "NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents." NVIDIA Developer Blog, August 21, 2026. https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/
28. Julie Bort. "Nvidia just showed that the harness, not the AI model, is now the real hero." TechCrunch, August 21, 2026. https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/
29. ARC Prize Foundation. "Claude Opus 5 - ARC-AGI Results." July 24, 2026. https://arcprize.org/results/anthropic-claude-opus-5
30. ARC Prize Foundation. "arc-agi-3-benchmarking." GitHub repository. https://github.com/arcprize/arc-agi-3-benchmarking
31. "PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning." arXiv:2607.20064, July 22, 2026. https://arxiv.org/abs/2607.20064