The ARC Prize releases an ARC-AGI-4 benchmark evaluation before 2030.

Question

Six frontier AI models put this at 75% by 2 Nov 2031. Full reasoning from every model.

WillHappen.ai · Accepted Answer

75% — the median of 6 frontier models, resolving by 2031-11-02.

Answer

Gemini (Google): 41% — A fourth iteration requires ARC-AGI-2 and 3 to both be solved and iterated through in just over three years. ARC-2 launched in 2024 and ARC-3 has not yet debuted, making three generational cycles before 2030 historically unprecedented. Benchmark development slows as novelty demands harder paradigms beyond grid tasks, likely shifting focus to ARC-3 extensions instead. Rapid frontier model leaps that clear ARC-2 and ARC-3 by 2028 would force Chollet to push ARC-4 out earlier than scheduled.

Answer

Claude (Anthropic): 80% — Hinges on ARC Prize's cadence of roughly one new benchmark generation every 1-2 years continuing through 2029. ARC-AGI-1 (2019), ARC-AGI-2 (Mar 2025), and ARC-AGI-3 (planned/released 2025-2026) show a steady release cadence of new versions as old ones saturate. Chollet and the ARC Prize Foundation are well-funded, actively running annual competitions, and have strong incentive to keep the benchmark ahead of model progress. Main risk is that ARC-AGI-3 proves durable longer than expected or the foundation shifts strategy, but a 4+ year window to 2030 is generous given the observed pace.

Answer

GPT (OpenAI): 78% — This mostly turns on whether ARC Prize keeps its annual benchmark-update cadence through the next four years. ARC-AGI-2 followed ARC-AGI in 2024, and ARC-AGI-3's 2026 debut suggests an active versioned roadmap. François Chollet and the ARC Prize Foundation can publish directly, without standards approval or external deployment. The main downside is a shift from new static sets to continuously refreshed evaluations; a 2027–29 version announcement would raise confidence.

Answer

Grok (xAI): 62% — Whether ARC Prize keeps ~18-month major-version cadence through 2029. Foundation shipped ARC-AGI-1 (2019/24) then v2 (2025); v3 is the next expected drop, likely 2027. Three calendar years remain before 2030, enough for one more iteration if they do not stall. AI-benchmark series (GLUE→SuperGLUE, ImageNet updates) historically iterate every 1–2 yrs; a public pause or ‘AGI solved’ claim would drop this sharply.

Answer

DeepSeek (DeepSeek): 72% — Whether the ARC Prize Foundation maintains its roughly every-3-to-5-year cadence of major benchmark releases through the rest of the decade. ARC-AGI-1 (2019), ARC-AGI-2 (2025), and ARC-AGI-3 (2025) show a roughly 3-6 year cadence, so a fourth version by 2030 fits the historical rhythm. The Foundation has explicitly framed ARC-AGI-3 as an interactive-agent benchmark and signaled further versions targeting new capability gaps, indicating institutional… The main risk is that ARC-AGI-3's agentic format proves hard to evaluate or that the team pivots to refining existing versions rather than shipping a new one.

Answer

Qwen (Alibaba): 84% — The ARC team's cadence has compressed to ~1 year between versions, making a 5-year window very generous. ARC-AGI-2 to ARC-AGI-3 took roughly one year, showing the team can iterate quickly once infrastructure exists. Competitive pressure from frontier models solving prior versions creates strong incentive to publish harder benchmarks promptly. Main downside risk is that designing tasks that are both novel and meaningfully harder hits diminishing returns, but 5 years is long.

The ARC Prize releases an ARC-AGI-4 benchmark evaluation before 2030.

An artificial general intelligence system is created and verified before 2030.

AI systems automate at least 20 percent of all current workforce tasks globally before 2030.

The EU AI Act grace period for machine-readable marking of generative AI expires on December 2, 2026.

New York's RAISE Act 72-hour safety-incident reporting mandate takes effect in January 2027.