lab ARC Prize · 597d ago

OpenAI o3 achieves breakthrough scores on ARC-AGI-Pub

from English

OpenAI's new o3 system, trained on the ARC-AGI-1 Public Training set, has scored 75.7% on the Semi-Private Evaluation set within a $10k compute limit and 87.5% with high-compute configuration.

The high-efficiency score qualifies as 1st place on the public leaderboard under ARC-AGI-Pub rules.
Testing was conducted at two compute levels: 6 samples for high-efficiency and 1024 samples (172x compute) for low-efficiency.
OpenAI confirmed the tested version differs from the officially released o3, with updated results to follow.

ARC Prize views this as a significant step-function increase in AI capabilities, demonstrating novel task adaptation abilities not seen in previous GPT-family models.

Importance 3/3 New feature vs. leaders ARC Prize OpenAI Benchmark results Reasoning models

Benchmarks

Benchmark	Model	Score
ARC-AGI 1	o3	75.7%

Read original

arxiv arXiv cs.CL · 52d ago

Frustrated Synchronization Network Outperforms Transformers

The Frustrated Synchronization Network (FSN) achieves lower validation loss than a RoPE-SwiGLU transformer at every epoch on character-level text and code tasks. At one million parameters, FSN converges to a validation loss of 1.5953 ± 0.0014, outperforming the transformer's converged loss of 1.611. This advantage persists up to four million parameters, with ongoing evaluations beyond that scale.

media AI Business Weekly · 137d ago · 3 views

OpenAI announces o3 retirement from ChatGPT and shares benchmark scores

OpenAI announced on May 28, 2026, that the o3 model will be retired from ChatGPT on August 26, 2026, following a 90-day sunset period. While the API remains available indefinitely, users are advised to migrate to successors like o4 or the GPT-5 family for complex reasoning and coding tasks.

media Jeremy Berman (Substack) · 198d ago · 7 views

GPT-5.2 Pro scores 31% on FrontierMath Tier 4

OpenAI's GPT-5.2 Pro achieved a new record score of 31% on the FrontierMath Tier 4 benchmark, solving 15 out of 48 problems. This represents a substantial improvement over the previous high score of 19% and demonstrates no evidence of over-fitting against OpenAI's exclusive access to 28 of the test problems.

lab OpenAI · 241d ago · 1 view

OpenAI releases GPT-5.2 Pro and Thinking models for science and math

OpenAI has released GPT-5.2 Pro and GPT-5.2 Thinking, describing them as its strongest models yet for scientific and mathematical work. The company aims to accelerate scientific research by helping researchers explore ideas and test hypotheses faster.

lab ARC Prize · 430d ago · 1 view

Testing major AI reasoning systems finds no clear winner on ARC benchmarks

A review of frontier AI reasoning systems from labs including OpenAI, DeepMind, xAI, DeepSeek, and Anthropic reveals that while test-time compute scaling improves accuracy, it significantly reduces efficiency. The analysis identifies a Pareto frontier for ARC-AGI-1 scores but notes complete failure on the more challenging ARC-AGI-2 benchmark, suggesting common limitations across all major models.