ARC-AGI-3
GPT-6 Astra 100% leads since Sep 2026 In use
Interactive game environments where agents must explore, infer goals and rules, and act efficiently without instructions.
Scores depend on the harness. ARC Prize reports GPT-6 Astra at 62.7 percent with its Standard harness and 99.9 percent with a Provider Adapter harness. A six-game preview ran from July 2025.
In the authors’ words
We introduce ARC-AGI-3, an interactive benchmark for studying agentic intelligence through novel, abstract, turn-based environments in which agents must explore, infer goals, build internal models of environment dynamics, and plan effective action sequences without explicit instructions. Like its predecessors ARC-AGI-1 and 2, ARC-AGI-3 focuses entirely on evaluating fluid adaptive efficiency on novel tasks, while avoiding language and external knowledge. ARC-AGI-3 environments only leverage Core Knowledge priors and are difficulty-calibrated via extensive testing with human test-takers. Our testing shows humans can solve 100% of the environments, in contrast to frontier AI systems which, as of March 2026, score below 1%. In this paper, we present the benchmark design, its efficiency-based scoring framework grounded in human action baselines, and the methodology used to construct, validate, and calibrate the environments.
Made by ARC Prize Foundation Released Mar 2026 Puzzles and chess Public tasks plus a private test set Last in a launch post Sep 2026
Last model added Oct 2026
Scores refreshed 2 Nov 2026
- Level
- Everyday Most adults, without training
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
24
Models scored on its own leaderboard
4
Launch posts list it, from 2 labs
1
System card reports it, from 1 lab
GPT-6 Astra leads with 100%
Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 100.0 GPT-6 Astra
- 97.5
Gemini 4 Argon - 96.4 GPT-6.1 Sol
- 72.4 Claude Sonnet 5.5
- 35.0 Gemini 3.8 Flash
- 30.2 Claude Opus 5
- 23.0 GPT-6 Sol
- 10.1 Grok 4.7
- 7.78 GPT-5.6 Sol
- 3.26
DeepSeek V4.1 Flash - 2.82 Claude Haiku 5.5
- 2.11 Grok 4.6
- 1.52 Claude Opus 4.8
- 0.80 GPT-5.6 Terra
- 0.59 GPT-6 Luna
- 0.51 Claude Opus 4.6
- 0.43 GPT-5.5
- 0.42 Gemini 3.1 Pro
- 0.42
Muse Spark Open - 0.32 Grok 4.5
- 0.21 GPT-5.4
- 0.18 Claude Opus 4.7
- 0.18 GPT-5.6 Luna
- 0.09 Grok 4.20
- 1 GPT-6 Astra high 100%
- 2 Gemini 4 Argon low 97.5%
- 3 GPT-6.1 Sol xhigh 96.4%
- 4 Claude Sonnet 5.5 xhigh 72.4%
- 5 Gemini 3.8 Flash high 35%
- 6 Claude Opus 5 high 30.2%
- 7 GPT-6 Sol max 23%
- 8 Grok 4.7 high 10.1%
- 9 GPT-5.6 Sol max 7.78%
- 10 DeepSeek V4.1 Flash low 3.26%
- 11 Claude Haiku 5.5 max 2.82%
- 12 Grok 4.6 xhigh 2.11%
- 13 Claude Opus 4.8 high 1.52%
- 14 GPT-5.6 Terra max 0.80%
- 15 GPT-6 Luna max 0.59%
- 16 Claude Opus 4.6 max 0.51%
- 17 GPT-5.5 high 0.43%
- 18 Gemini 3.1 Pro 0.42%
- 19 Muse Spark Open xhigh 0.42%
- 20 Grok 4.5 medium 0.32%
- 21 GPT-5.4 high 0.21%
- 22 Claude Opus 4.7 high 0.18%
- 23 GPT-5.6 Luna max 0.18%
- 24 Grok 4.20 beta 0.09%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 100% | high, Provider Adapter | Sep 2026 | Official leaderboard |
| 2 | Gemini 4 Argon | 97.5% | low, Provider Adapter | Sep 2026 | Official leaderboard |
| 3 | GPT-6.1 Sol | 96.4% | xhigh, Provider Adapter | Sep 2026 | Official leaderboard |
| 4 | Claude Sonnet 5.5 | 72.4% | xhigh, Provider Adapter | Sep 2026 | Official leaderboard |
| 5 | Gemini 3.8 Flash | 35% | high, Provider Adapter | Sep 2026 | Official leaderboard |
| 6 | Claude Opus 5 | 30.2% | high, standard | Jul 2026 | Official leaderboard |
| 7 | GPT-6 Sol | 23% | max, Provider Adapter | Sep 2026 | Official leaderboard |
| 8 | Grok 4.7 | 10.1% | high, Provider Adapter | Sep 2026 | Official leaderboard |
| 9 | GPT-5.6 Sol | 7.78% | max, standard | Jul 2026 | Official leaderboard |
| 10 | DeepSeek V4.1 Flash | 3.26% | low, Provider Adapter | Sep 2026 | Official leaderboard |
| 11 | Claude Haiku 5.5 | 2.82% | max, Provider Adapter | Oct 2026 | Official leaderboard |
| 12 | Grok 4.6 | 2.11% | xhigh, standard | Aug 2026 | Official leaderboard |
| 13 | Claude Opus 4.8 | 1.52% | high, standard | May 2026 | Official leaderboard |
| 14 | GPT-5.6 Terra | 0.80% | max, standard | Jul 2026 | Official leaderboard |
| 15 | GPT-6 Luna | 0.59% | max, Provider Adapter | Sep 2026 | Official leaderboard |
| 16 | Claude Opus 4.6 | 0.51% | max, standard | Feb 2026 | Official leaderboard |
| 17 | GPT-5.5 | 0.43% | high, standard | Apr 2026 | Official leaderboard |
| 18 | Gemini 3.1 Pro | 0.42% | standard | Feb 2026 | Official leaderboard |
| 19 | Muse Spark Open | 0.42% | xhigh, Provider Adapter | Oct 2026 | Official leaderboard |
| 20 | Grok 4.5 | 0.32% | medium, standard | Jul 2026 | Official leaderboard |
| 21 | GPT-5.4 | 0.21% | high, standard | Mar 2026 | Official leaderboard |
| 22 | Claude Opus 4.7 | 0.18% | high, standard | Apr 2026 | Official leaderboard |
| 23 | GPT-5.6 Luna | 0.18% | max, standard | Jul 2026 | Official leaderboard |
| 24 | Grok 4.20 | 0.09% | beta, standard | Feb 2026 | Official leaderboard |
Labs report scores for 4 models in their cards. We plot every source in the charts below.
More effort adds a median 4 points
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Lowest effort to highest, light to dark
GPT-6 Astra Provider Adapter
none
100% Gemini 4 Argon Provider Adapter low
97.5% GPT-6.1 Sol Provider Adapter
low medium
96.4% Claude Sonnet 5.5 Provider Adapter
low high xhigh
72.4% GPT-6 Astra standard
none low medium high max
62.7% GPT-6.1 Sol standard
low high xhigh max
52.7% Gemini 4 Argon standard low
49.8% Gemini 3.8 Flash Provider Adapter low medium high
35% Claude Sonnet 5.5 standard
low xhigh
24.3% GPT-6 Sol Provider Adapter
none xhigh max
23% Gemini 3.8 Flash standard low
10.4% Grok 4.7 Provider Adapter
low medium
10.1% | Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-6 Astra | high | Provider Adapter | 100% | ||
| GPT-6 Astra | max | Provider Adapter | 98.5% | ||
| GPT-6 Astra | xhigh | Provider Adapter | 98.4% | ||
| GPT-6 Astra | medium | Provider Adapter | 98.4% | ||
| GPT-6 Astra | low | Provider Adapter | 98% | ||
| Gemini 4 Argon | low | Provider Adapter | 97.5% | ||
| Gemini 4 Argon | medium | Provider Adapter | 97.3% | ||
| Gemini 4 Argon | high | Provider Adapter | 97.3% | ||
| GPT-6 Astra | none | Provider Adapter | 96.7% | ||
| GPT-6.1 Sol | xhigh | Provider Adapter | 96.4% | ||
| GPT-6.1 Sol | max | Provider Adapter | 96.2% | ||
| GPT-6.1 Sol | high | Provider Adapter | 95% | ||
| GPT-6.1 Sol | medium | Provider Adapter | 91% | ||
| GPT-6.1 Sol | low | Provider Adapter | 82.8% | ||
| Claude Sonnet 5.5 | xhigh | Provider Adapter | 72.4% | ||
| GPT-6 Astra | max | standard | 62.7% | ||
| GPT-6 Astra | xhigh | standard | 59.3% | ||
| GPT-6 Astra | high | standard | 54.8% | ||
| GPT-6.1 Sol | max | standard | 52.7% | ||
| Gemini 4 Argon | low | standard | 49.8% | ||
| Gemini 4 Argon | medium | standard | 46.3% | ||
| GPT-6.1 Sol | xhigh | standard | 39.9% | ||
| GPT-6 Astra | medium | standard | 38.6% | ||
| Claude Sonnet 5.5 | high | Provider Adapter | 35.5% | ||
| GPT-6 Astra | none | standard | 35.2% | ||
| Gemini 3.8 Flash | high | Provider Adapter | 35% | ||
| Claude Opus 5 | high | standard | 30.2% | ||
| GPT-6.1 Sol | high | standard | 26.7% | ||
| Claude Sonnet 5.5 | xhigh | standard | 24.3% | ||
| Gemini 3.8 Flash | medium | Provider Adapter | 24.2% | ||
| GPT-6 Sol | max | Provider Adapter | 23% | ||
| GPT-6 Astra | low | standard | 17.4% | ||
| Gemini 3.8 Flash | low | Provider Adapter | 15.1% | ||
| GPT-6.1 Sol | medium | standard | 10.6% | ||
| Gemini 3.8 Flash | high | standard | 10.4% | ||
| Grok 4.7 | high | Provider Adapter | 10.1% | ||
| GPT-6 Sol | xhigh | Provider Adapter | 9.45% | ||
| Grok 4.7 | medium | Provider Adapter | 8.37% | ||
| GPT-5.6 Sol | max | standard | 7.78% | ||
| Grok 4.7 | xhigh | Provider Adapter | 7.02% | ||
| GPT-5.6 Sol | xhigh | standard | 6.99% | ||
| Gemini 3.8 Flash | low | standard | 5.99% | ||
| GPT-6 Sol | high | Provider Adapter | 5.86% | ||
| GPT-6 Sol | medium | Provider Adapter | 4.93% | ||
| Claude Sonnet 5.5 | medium | Provider Adapter | 4.89% | ||
| GPT-6 Sol | max | standard | 4.62% | ||
| Gemini 3.8 Flash | medium | standard | 3.99% | ||
| GPT-6.1 Sol | low | standard | 3.92% | ||
| DeepSeek V4.1 Flash | low | Provider Adapter | 3.26% | ||
| DeepSeek V4.1 Flash | high | Provider Adapter | 3.15% | ||
| Claude Haiku 5.5 | max | Provider Adapter | 2.82% | ||
| DeepSeek V4.1 Flash | max | Provider Adapter | 2.68% | ||
| GPT-5.6 Sol | high | standard | 2.15% | ||
| Grok 4.6 | xhigh | standard | 2.11% | ||
| Grok 4.7 | xhigh | standard | 1.82% | ||
| GPT-6 Sol | xhigh | standard | 1.80% | ||
| GPT-6 Sol | low | Provider Adapter | 1.76% | ||
| Grok 4.7 | high | standard | 1.73% | ||
| Grok 4.7 | medium | standard | 1.71% | ||
| Claude Opus 4.8 | high | standard | 1.52% | ||
| GPT-6 Sol | none | Provider Adapter | 1.21% | ||
| Claude Sonnet 5.5 | high | standard | 1.20% | ||
| GPT-5.6 Sol | medium | standard | 1.07% | ||
| GPT-6 Sol | high | standard | 0.90% | ||
| DeepSeek V4.1 Flash | max | standard | 0.85% | ||
| GPT-5.6 Terra | max | standard | 0.80% | ||
| Grok 4.7 | low | Provider Adapter | 0.77% | ||
| GPT-5.6 Terra | xhigh | standard | 0.65% | ||
| Claude Sonnet 5.5 | low | Provider Adapter | 0.62% | ||
| GPT-6 Luna | max | Provider Adapter | 0.59% | ||
| Claude Opus 4.6 | max | standard | 0.51% | ||
| GPT-6 Luna | xhigh | Provider Adapter | 0.51% | ||
| GPT-5.6 Terra | high | standard | 0.49% | ||
| DeepSeek V4.1 Flash | high | standard | 0.48% | ||
| DeepSeek V4.1 Flash | none | Provider Adapter | 0.48% | ||
| Claude Haiku 5.5 | max | standard | 0.45% | ||
| GPT-5.5 | high | standard | 0.43% | ||
| Gemini 3.1 Pro | standard | 0.42% | |||
| GPT-6 Sol | medium | standard | 0.42% | ||
| Muse Spark Open | xhigh | Provider Adapter | 0.42% | ||
| Claude Sonnet 5.5 | low | standard | 0.40% | ||
| Claude Sonnet 5.5 | medium | standard | 0.40% | ||
| GPT-6 Luna | high | Provider Adapter | 0.38% | ||
| DeepSeek V4.1 Flash | low | standard | 0.35% | ||
| DeepSeek V4.1 Flash | none | standard | 0.34% | ||
| GPT-5.6 Sol | low | standard | 0.33% | ||
| Grok 4.5 | medium | standard | 0.32% | ||
| GPT-6 Luna | medium | Provider Adapter | 0.32% | ||
| Grok 4.5 | high | standard | 0.30% | ||
| GPT-6 Sol | none | standard | 0.28% | ||
| Grok 4.5 | low | standard | 0.26% | ||
| Grok 4.7 | low | standard | 0.25% | ||
| GPT-6 Luna | low | Provider Adapter | 0.24% | ||
| GPT-5.4 | high | standard | 0.21% | ||
| GPT-6 Luna | medium | standard | 0.19% | ||
| GPT-6 Luna | high | standard | 0.18% | ||
| Claude Opus 4.7 | high | standard | 0.18% | ||
| GPT-5.6 Luna | max | standard | 0.18% | ||
| GPT-5.6 Luna | low | standard | 0.17% | ||
| GPT-5.6 Luna | medium | standard | 0.17% | ||
| GPT-6 Luna | xhigh | standard | 0.16% | ||
| Muse Spark Open | xhigh | standard | 0.14% | ||
| GPT-6 Sol | low | standard | 0.13% | ||
| GPT-6 Luna | max | standard | 0.10% | ||
| GPT-5.6 Luna | high | standard | 0.10% | ||
| Grok 4.20 | beta | standard | 0.09% | ||
| GPT-5.6 Terra | medium | standard | 0.08% | ||
| GPT-6 Luna | none | Provider Adapter | 0.05% | ||
| GPT-6 Luna | low | standard | 0.03% | ||
| GPT-6 Luna | none | standard | 0.03% | ||
| GPT-5.6 Luna | xhigh | standard | 0.02% | ||
| GPT-5.6 Terra | low | standard | 0.01% |
6 models have held number one since its release in Mar 2026
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
Anthropic OpenAI Google DeepMind
The best score passed 85% in Sep 2026
All the scores we have, one dot each, so a model tested at several settings appears more than once.
On the benchmark’s own leaderboard Reported by the lab in its model card
Best so far
2 labs put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 3 Sep 2026 | GPT-6 Astra | 99.9 | ARC-AGI-3 (responses API harness) |
| 24 Jul 2026 | Claude Opus 5 | 30.2 | Novel problem-solving ARC-AGI-3 (high) |
| 9 Jul 2026 | GPT-5.6 Luna | 0.18 | ARC-AGI-3 |
| 9 Jul 2026 | GPT-5.6 Terra | 0.8 | ARC-AGI-3 |
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 24 Jul 2026 | Claude Opus 5 system card | 30.2 | ARC-AGI-3 semi-private set, RHAE score 30.16% at high effort, max effort not available at release, verified by ARC Prize Foundation (Table 8.1.A, 8.14.2) | Capabilities |
Versions
We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.
OpenAI ByteDance Seed xAI Anthropic Google DeepMind
More in Games and long tasks
- MineDojo Fading Agents complete open-ended tasks in Minecraft, such as mining blocks and fighting mobs, from a large simulation suite. Programmatic checks score many tasks.
Seed2.0 Pro 49% - Orak Fading LLM agents play 12 popular video games across most genres through an MCP interface. Scores come from in-game progress or outcomes per title.
Nemotron 3 Nano 4B 22.9 - GameWorld Only in system cards Multimodal game agents play 34 browser games across 170 tasks. Outcomes are checked from the game state. Gemini 3 Flash 41.9%
- Swarm Bench Only in system cards Models orchestrate agent swarms on complex tasks that benefit from coordinated decomposition and parallel execution. Kimi K3 76.3%
- TextQuests Only in system cards Agents play 25 classic interactive fiction games. Measures game progress and harmful actions per 100 steps. Muse Spark 2.40
- Anthropic large agent teams tasks Only in system cards Internal 24-hour tasks for teams of 1-100 agents: build a company knowledge base, or formalize a theorem in Lean.