ARC-AGI-2 best score
Running best ARC-AGI-2 score by release date, from Epoch's compilation of external results. Derived metric arc_agi_frontier.
Reason through thisThe reading
External, compute-unconstrained results; the score says nothing about cost per task, which the prize also caps.
17 readings drawn from the derived metric arc_agi_frontier; 17 observations in the evidence.
Derived (arc_agi_frontier) from: Epoch AI benchmarks and Capabilities Index. Epoch AI, Benchmarking Hub and Epoch Capabilities Index (CC BY 4.0).
The numbers
| as of | reading | value | record |
|---|---|---|---|
| 2026-09-03 | — | 95.0% | reported |
| 2026-07-09 | — | 92.5% | reported |
| 2026-06-09 | — | 89.2% | reported |
| 2026-04-23 | — | 85.0% | reported |
| 2026-02-12 | — | 84.6% | reported |
| 2026-02-05 | — | 69.2% | reported |
| 2025-12-11 | — | 54.2% | reported |
| 2025-11-24 | — | 37.6% | reported |
| 2025-11-18 | — | 31.1% | reported |
| 2025-10-07 | — | 18.3% | reported |
| 2025-07-09 | — | 16.0% | reported |
| 2025-05-22 | — | 8.6% | reported |
| 2025-04-16 | — | 6.5% | reported |
| 2025-01-31 | — | 3.0% | reported |
| 2025-01-20 | — | 1.3% | reported |
| 2024-09-12 | — | 0.8% | reported |
| 2024-07-18 | — | 0.0% | reported |
Why it matters
The one serious proxy for out-of-distribution generalisation, the brief's replacement for a 'world-model transfer' dimension.
What cuts against it
Compiled from external submissions rather than run by Epoch; cost per task is unconstrained in most entries; a single benchmark.
How it is measured
- derived metric
arc_agi_frontier(formula in the semantic layer) - source Epoch AI benchmarks and Capabilities Index · default tier 1 · CC BY 4.0
- proxy benchmark · cadence monthly · valve invention to product
Normal = under half (the 2025 frontier); fast = 85% or above (the ARC Prize grand-prize bar). Between is emerging.
Applied to metric:arc_agi_frontier.
Derived rows (17)
| as of | dims | value | inputs |
|---|---|---|---|
| 2026-09-03 | 95.0% | source1 | |
| 2026-07-09 | 92.5% | source1 | |
| 2026-06-09 | 89.2% | source1 | |
| 2026-04-23 | 85.0% | source1 | |
| 2026-02-12 | 84.6% | source1 | |
| 2026-02-05 | 69.2% | source1 | |
| 2025-12-11 | 54.2% | source1 | |
| 2025-11-24 | 37.6% | source1 | |
| 2025-11-18 | 31.1% | source1 | |
| 2025-10-07 | 18.3% | source1 | |
| 2025-07-09 | 16.0% | source1 | |
| 2025-05-22 | 8.6% | source1 | |
| 2025-04-16 | 6.5% | source1 | |
| 2025-01-31 | 3.0% | source1 | |
| 2025-01-20 | 1.3% | source1 | |
| 2024-09-12 | 0.8% | source1 | |
| 2024-07-18 | 0.0% | source1 |
Latest readings
Why this status, and its history
Epoch's compilation of external ARC-AGI-2 results: the running best score reached 0.95 with GPT-6 Astra (3 Sep 2026), past the fast band (85%, the ARC Prize bar), from 0.89 in June 2026. Compute-unconstrained submissions compiled from a leaderboard are tier 6 and single-source, so the status is capped at emerging. Initial seed.
- 2026-09-10unmeasured to emergingconf — → 55 · the evaluator (automatic)
Epoch's compilation of external ARC-AGI-2 results: the running best score reached 0.95 with GPT-6 Astra (3 Sep 2026), past the fast band (85%, the ARC Prize bar), from 0.89 in June 2026. Compute-unconstrained submissions compiled from a leaderboard are tier 6 and single-source, so the status is capped at emerging. Initial seed.
sources123456+197 more
Why this confidence
Rests on 17 observations from Epoch AI benchmarks and Capabilities Index; best evidence tier 6, grade B; current.
Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.
Predictions that lean on this
- Happening
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
More predictions that lean on this
- Too early to tell
- Too early to tell