OY Labs Achieves A Leading ARC-AGI-3 Public Score In Its Mission To Build The World's Smartest AI

(MENAFN- GlobeNewsWire - Nasdaq) OY1-AGI reaches 100% at a recorded cost of $415.37, exceeding OpenAI's High-reasoning public results and NVIDIA NOOA's listed score

WARSAW, Poland, Sept. 11, 2026 (GLOBE NEWSWIRE) -- AI research company OY Labs today announced a perfect 100.0% ARC-AGI-3 public benchmark score for OY1-AGI, winning all 25 games and solving all 183 levels. Using GPT-6 Astra at High reasoning, the system achieved 100% in every environment, exceeding the public results published by ARC Prize for the same model and reasoning setting with the Provider Adapter harness.

The main result, published on 9 September, used 6,732 actions. OY Labs records a total API cost of $415.37, normal completion in 6 hours 48 minutes, and 37,704 evidence consistency checks passed. Three public scorecards now document perfect scores, with links to every run on the following page.

The result reaches the public scoring ceiling and the highest score currently shown on the ARC-AGI-3 Community Leaderboard. OY Labs says its application is under review; the three scorecards are already publicly available.

OpenAI and NVIDIA benchmark comparisons

On the public set, ARC Prize reports 100% in 24 environments and 91.8% in TN36 for GPT-6 Astra with the Provider Adapter at High reasoning. OY1-AGI achieved 100% in all 25. Separately, ARC Prize's verified semi-private result for that OpenAI configuration is 99.95% at $18,817. OY1's $415.37 public-run cost is substantially lower, although these cost figures cover different evaluation sets.

System and configuration	Test set		Score		Run cost
OY Labs OY1-AGI	Public		100.0%		$415.37
GPT-6 Astra · High reasoning
OpenAI GPT-6 Astra	Semi-private		99.95%		$18,817
Provider Adapter · High reasoning
Retrodict	Public		99.9%		$654
NVIDIA NOOA	Public		85.1%		$332

^{Selected results checked on 10 September 2026. Public and semi-private evaluations are not directly comparable; systems and cost accounting also differ. OY1's leaderboard application remains under review.}

OY1-AGI also exceeds Retrodict's 99.9%, the next-highest community score below 100%, and NVIDIA NOOA's 85.1%. The NVIDIA comparison concerns different systems, rather than the same model and reasoning configuration.

Three publicly available OY1-AGI scorecards

All three scorecards are linked below. The 9 September result is the main run highlighted in this release, with a recorded API cost of $415.37.

Building and training our own large language models

OY Labs' initial priority is developing and training its own large language models for mainstream AI users and institutions, including models trained to meet institutions' specific requirements.

OY Labs reports tremendous interest in its product as it pursues leading AI capability at the lowest possible cost per unit of intelligence.

“Our goal is to make advanced intelligence affordable enough to use everywhere,” said Laurin Bylica, Cofounder and CEO of OY Labs.

OY2-AGI built on ethical values from the outset

A core differentiator for OY2-AGI will be its foundation in strong ethical values from the outset. OY Labs plans to embed a hard-coded constitution into the model's design, with safeguards intended to protect user privacy and prevent harmful use.

“We are building powerful AI models to advance human welfare. That purpose should shape both the capabilities we develop and the boundaries we build into them,” said Loïc Bellez, Cofounder and CTO of OY Labs.

The current benchmark system

The benchmark result was produced by GPT-6 Astra at High reasoning with OY Labs' OY1 harness, which guides the model's problem-solving process.

OY Labs has open-sourced the benchmark implementation under the Apache 2.0 license, including the evaluated source, reproduction instructions and results documentation. Explore the OY1 benchmark repository.

About OY Labs

OY Labs is an AI research company founded by Laurin Bylica and Loïc Bellez and operated by Orca Labs sp. z o.o., registered in Poland.

Sources and notes to editors

The scorecards substantiate scores, levels, environments and action counts. OY Labs reports cost, runtime, normal completion and 37,704 evidence consistency checks; these software checks are not ARC Prize certification. The repository states that public games were used during development. These results do not establish a held-out ranking or general superiority across AI tasks. Product interest and application-review status are company statements.

9 September · Main:

7 September:

6 September:

Benchmark comparisons: ARC Prize's OpenAI results · Community Leaderboard

Media contact: ... · oylabs.ai

OY Labs Contact Details:
Orca Labs sp. z o.o.
ul. Marcina Kasprzaka 31/119
01-234 Warsaw
Poland

Published scorecard	Score		Games won		Levels solved		Actions
9 September 2026 · Main result	100.0%		25 / 25		183 / 183		6,732
7 September 2026	100.0%		25 / 25		183 / 183		6,714
6 September 2026	100.0%		25 / 25		183 / 183		6,659

MENAFN11092026004107003653ID1111650646