Benchmark · reasoning

ARC-AGI 3

12 results 7 models

ARC Prize's third generation (2026): the agent must learn an unseen interactive game environment from scratch. Scores run far lower than — and are NOT comparable to — the static grid-puzzle ARC-AGI 1/2.

Schema Claude Opus 5 GPT‑5.6 Sol Prime Agent GPT-6 Astra Nvidia AVO GPT‑6 Astra

Timeline

Date	Model	Score	Source
2026-09-07	GPT‑6 Astra	99.9%	Claude proves Fermat's Last Theorem, OpenAI plans automated researcher
2026-09-05	GPT‑6 Astra	99.9%	OpenAI releases GPT-6 Astra with improved computer use, coding, and cybersecurity capabilities
2026-09-04	GPT-6 Astra	62.7%	Nvidia acquires Hugging Face; OpenAI releases GPT-6 Astra; Microsoft launches MAI-Transcribe-2
2026-09-04	Nvidia AVO	100.0%	Nvidia demonstrates 100% on ARC AGI-3 with AVO harness
2026-09-04	GPT-6 Astra	99.9%	OpenAI releases GPT-6 Astra, a computer-use model with 1.05M context
2026-08-25	Prime Agent	95.5%	Prime Agent: A self-improving RLM harness raises ARC-AGI-3 RHAE Best@1 to 95.5%
2026-08-25	Prime Agent	95.5%	Prime Agent: A self-improving RLM harness raises ARC-AGI-3 RHAE Best@1 to 95.5%
2026-08-13	GPT‑5.6 Sol	38.3%	OpenAI releases GPT-5.6 for affordable, high-performance agentic workflows
2026-08-06	Prime Agent	95.5%	PrimeIntellect releases Prime Agent, a self-improving coding harness
2026-07-31	GPT‑5.6 Sol	38.3%	OpenAI reduces GPT-5.6 Luna price by 80 percent
2026-07-31	Claude Opus 5	30.2%	Anthropic launches Claude Opus 5; OpenAI models breach Hugging Face during security test
2026-07-17	Schema	99.0%	Moonshot launches Kimi K3; Schema harness crushes ARC-AGI 3