ReasoningOfficial + Vendor scoresPrices verified
ARC-AGI-3 leaderboard
As of November 8, 2026, GPT-6 Astra (OpenAI) leads ARC-AGI-3 at 62.7%, ahead of GPT-6.1 Sol at 52.7% and Gemini 4 Argon at 49.8%. 24 systems ranked. All 24 scores on the main board come from ARC Prize Foundation, not from the labs that make the models.
Top score
62.7%
Systems ranked
24
Independent rows
35/37
Cheapest near top
$6.82/1M
With API prices
22/24
Newest score
Oct 17
What is ARC-AGI-3? ARC-AGI-3 is an interactive reasoning benchmark: the agent plays turn-based grid games with no instructions and has to work out the rules and the goal by exploring. Made by ARC Prize Foundation, March 2026.
§ 01 / LEADERBOARD
Top models on ARC-AGI-3, November 2026
ARC Prize's minimal, provider-neutral interface: the model keeps its own notes.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | ARC Prize verified, best result | $10 / $50 | 1.05M | Official leaderboard ARC Prize Foundation |
| 2 | GPT-6.1 Sol OpenAI | ARC Prize verified, best result | $2 / $10 | 1.05M | Official leaderboard ARC Prize Foundation |
| 3 | Gemini 4 Argon Google DeepMind | ARC Prize verified, best result | $2 / $10 | 1M | Official leaderboard ARC Prize Foundation |
| 4 | Claude Opus 5 Anthropic | ARC Prize verified, best result | $5 / $25 | 1M | Official leaderboard ARC Prize Foundation |
| 5 | Claude Sonnet 5.5 Anthropic | ARC Prize verified, best result | $2 / $10 | 1M | Official leaderboard ARC Prize Foundation |
| 6 | Gemini 3.8 Flash Google DeepMind | ARC Prize verified, best result | $0.75 / $3.75 | 1M | Official leaderboard ARC Prize Foundation |
| 7 | GPT-5.6 Sol OpenAI | ARC Prize verified, best result | $4 / $20 | 1.05M | Official leaderboard ARC Prize Foundation |
| 8 | GPT-6 Sol OpenAI | ARC Prize verified, best result | $2 / $10 | 1.05M | Official leaderboard ARC Prize Foundation |
| 9 | Grok 4.6 SpaceXSI | ARC Prize verified, best result | $2 / $6 | 500K | Official leaderboard ARC Prize Foundation |
| 10 | Grok 4.7 SpaceXSI | ARC Prize verified, best result | $2 / $6 | 500K | Official leaderboard ARC Prize Foundation |
| 11 | Claude Opus 4.8 Anthropic | ARC Prize verified, best result | $5 / $25 | 1M | Official leaderboard ARC Prize Foundation |
| 12 | GPT-5.6 Terra OpenAI | ARC Prize verified, best result | $2 / $12 | 1.05M | Official leaderboard ARC Prize Foundation |
| 13 | Claude Haiku 5.5 Anthropic | ARC Prize verified, best result | $0.1 / $0.5 | 1M | Official leaderboard ARC Prize Foundation |
| 13 | Claude Opus 4.6 Anthropic · listed as “Anthropic Opus 4.6” | ARC Prize verified, best result | $5 / $25 | 1M | Official leaderboard ARC Prize Foundation |
| 15 | Gemini 3.1 Pro Preview Preview Google | ARC Prize verified, best result | $2 / $12 | 1M | Official leaderboard ARC Prize Foundation |
| 15 | GPT-5.5 OpenAI | ARC Prize verified, best result | $5 / $30 | 1M | Official leaderboard ARC Prize Foundation |
| 17 | Grok 4.5 SpaceXSI | ARC Prize verified, best result | $2 / $6 | 500K | Official leaderboard ARC Prize Foundation |
| 18 | Claude Opus 4.7 Anthropic · listed as “Opus 4.7” | ARC Prize verified, best result | $5 / $25 | 1M | Official leaderboard ARC Prize Foundation |
| 18 | DeepSeek V4.1 Flash DeepSeek · listed as “DeepSeek V4.1 Flash 1017” | ARC Prize verified, best result | $0.3 / $1.2 | 1M | Official leaderboard ARC Prize Foundation |
| 18 | GPT-5.4 OpenAI | ARC Prize verified, best result | $2.5 / $15 | 1.05M | Official leaderboard ARC Prize Foundation |
| 18 | GPT-5.6 Luna OpenAI | ARC Prize verified, best result | $0.2 / $1.2 | 1.05M | Official leaderboard ARC Prize Foundation |
| 18 | GPT-6 Luna OpenAI | ARC Prize verified, best result | $0.1 / $0.5 | 1.05M | Official leaderboard ARC Prize Foundation |
| 23 | Grok 4.20 (Beta Reasoning) Not in our price index | ARC Prize verified, best result | — | — | Official leaderboard ARC Prize Foundation |
| 23 | Muse Spark Open Not in our price index | ARC Prize verified, best result | — | — | Official leaderboard ARC Prize Foundation |
§ 02 / OTHER SETTINGS
ARC-AGI-3 under other settings
Labs report ARC-AGI-3 under different settings, and a score from one setting is not comparable with a score from another. Each setting has its own table, ranked on its own.
02.1
Provider adapter
The provider's own context management (reasoning state kept between turns, compaction).
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | ARC Prize verified, best result | $10 / $50 | 1.05M | Official leaderboard ARC Prize Foundation |
| 2 | Gemini 4 Argon Google DeepMind | ARC Prize verified, best result | $2 / $10 | 1M | Official leaderboard ARC Prize Foundation |
| 3 | GPT-6.1 Sol OpenAI | ARC Prize verified, best result | $2 / $10 | 1.05M | Official leaderboard ARC Prize Foundation |
| 4 | Claude Sonnet 5.5 Anthropic | ARC Prize verified, best result | $2 / $10 | 1M | Official leaderboard ARC Prize Foundation |
| 5 | Gemini 3.8 Flash Google DeepMind | ARC Prize verified, best result | $0.75 / $3.75 | 1M | Official leaderboard ARC Prize Foundation |
| 6 | GPT-6 Sol OpenAI | ARC Prize verified, best result | $2 / $10 | 1.05M | Official leaderboard ARC Prize Foundation |
| 7 | Grok 4.7 SpaceXSI | ARC Prize verified, best result | $2 / $6 | 500K | Official leaderboard ARC Prize Foundation |
| 8 | DeepSeek V4.1 Flash DeepSeek · listed as “DeepSeek V4.1 Flash 1017” | ARC Prize verified, best result | $0.3 / $1.2 | 1M | Official leaderboard ARC Prize Foundation |
| 9 | Claude Haiku 5.5 Anthropic | ARC Prize verified, best result | $0.1 / $0.5 | 1M | Official leaderboard ARC Prize Foundation |
| 10 | GPT-6 Luna OpenAI | ARC Prize verified, best result | $0.1 / $0.5 | 1.05M | Official leaderboard ARC Prize Foundation |
| 11 | Muse Spark Open Not in our price index | ARC Prize verified, best result | — | — | Official leaderboard ARC Prize Foundation |
02.2
Harness not stated
Lab-reported numbers that do not say which harness they used.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | as reported | $10 / $50 | 1.05M | Vendor-reported OpenAI |
| 2 | GPT-5.6 Sol OpenAI | as reported | $4 / $20 | 1.05M | Vendor-reported OpenAI |
§ 03 / SCORE VS PRICE
ARC-AGI-3: score against price
§ 04 / PROVENANCE
Can you trust ARC-AGI-3 scores?
35 scores on this page come from the benchmark's own leaderboard (ARC Prize Foundation), where every model runs on the same harness. The other 2 are labs' reports on their own models.
- ARC Prize publishes two numbers per model. The standard harness is a minimal, provider-neutral interface where the model keeps its own notes; the provider adapter uses the provider's own context management. GPT-6 Astra scored 62.7% on the first and 99.9% on the second, so the two are separate tables here.
- The semi-private set has been exposed to API providers during testing, so limited leakage is possible.
- A high score needs both completion and few actions; a model that finishes levels slowly still scores low.
§ 05 / HOW IT WORKS
How ARC-AGI-3 works
The task
The agent acts in a 2D grid environment with no stated goal. It has to explore, discover what the controls do and what winning means, and clear increasingly hard levels.
The score
Action efficiency against people: each level scores by how few actions the agent needed compared with a human baseline, capped, then averaged per game and across games.
Test set
The authors do not publish a single task count for this version.
Made by
ARC Prize Foundation, first published March 2026.
Official sources
- arcprize.org/arc-agi/3
- Definition read on 2026-11-08
§ 06 / SOURCES
Sources and freshness
37 published scores from 2 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.
- Official leaderboard ARC Prize Foundation
- Vendor-reported OpenAI
ARC-AGI-3 FAQ
Answers derived from the 37 published scores on this page and the current price index.
Q · 01 Which AI model has the highest ARC-AGI-3 score?
+
GPT-6 Astra with 62.7% (ARC Prize verified, best result), on ARC Prize Foundation's own leaderboard on September 2, 2026. Next: GPT-6.1 Sol at 52.7% and Gemini 4 Argon at 49.8%.
Q · 02 What is the cheapest model near the top of ARC-AGI-3?
+
The leader itself. No priced model within 5 points of GPT-6 Astra is cheaper: it runs at $6.82 per 1M tokens on our effective blend ($10 input, $50 output).
Q · 03 Are these ARC-AGI-3 scores independently verified?
+
35 of 37 published scores come from ARC Prize Foundation; the other 2 are vendor reports. Each row carries a badge that says which, and a link to the source.
Q · 04 What does ARC-AGI-3 measure?
+
ARC-AGI-3 is an interactive reasoning benchmark: the agent plays turn-based grid games with no instructions and has to work out the rules and the goal by exploring. The agent acts in a 2D grid environment with no stated goal. It has to explore, discover what the controls do and what winning means, and clear increasingly hard levels.
Q · 05 How is the ARC-AGI-3 score calculated?
+
Action efficiency against people: each level scores by how few actions the agent needed compared with a human baseline, capped, then averaged per game and across games.
Q · 06 How current is this ARC-AGI-3 leaderboard?
+
The newest score was published on October 17, 2026. Every row shows the date its source published it, and we last read all of them on November 8, 2026. Prices come from our model price index, where each rate carries the date it was read on the vendor's page.