Tufa Labs Wins $25K Beating Frontier AI on the World's Hardest Benchmark

The hardest AI benchmark alive just handed out its first prize money. ARC Prize 2026 has awarded its Milestone #1 prizes for the ARC-AGI-3 competition, recognizing the top three teams who open-sourced their agents on the Kaggle leaderboard by the June 30 deadline. The winners: Tufa Labs in first, Reki in second, and Md Boktiar Mahbub Murad in third. All three notebooks are now public.

A benchmark that breaks everything frontier AI has

ARC-AGI-3 is a fundamentally different kind of benchmark. It comprises hundreds of original turn-based environments, each handcrafted by human game designers, with no instructions, no rules, and no stated goals. To succeed, an AI agent must explore each environment on its own, figure out how it works, discover what winning looks like, and carry what it learns forward across increasingly difficult levels.

This marks the first major format change since ARC-AGI was introduced in 2019. While the first two versions challenged static reasoning, the third version is designed to challenge interactive reasoning and requires new AI capabilities: exploration, planning, memory, goal acquisition, and alignment.

The gap between human and AI performance is staggering. Humans can solve 100% of the environments, in contrast to frontier AI systems which, as of March 2026, score below 1%. To put that in context: by early 2026, Gemini 3.1 Pro was scoring 98% on ARC-AGI-1. That benchmark was effectively solved. ARC-AGI-3 reset the clock entirely.

The scoring system is also novel. Due to the operational intensity of running a full evaluation using high-reasoning frontier model APIs, there is a hard cutoff of 5x human performance per level. If a human takes 10 actions to beat a certain level on average, the AI agent is cut off after 50 actions. Efficiency is the metric, not just completion.

The milestone prize structure

Participants who open-source their solutions by the milestone deadlines are eligible for milestone prize money. Milestone #1 (June 30, 2026): 1st place receives $25K, 2nd receives $10K, and 3rd receives $2.5K. A second milestone checkpoint follows on September 30 with the same payout structure.

The three winning notebooks and their scores:

• Tufa Labs (1st place) , Score: 1.03 on the Kaggle leaderboard, submitted as "duck harness"
• Reki (2nd place) , Milestone #1 second-place solution
• Md Boktiar Mahbub Murad (3rd place) , Score: 0.86 on the leaderboard

ARC Prize 2026 has a total prize pool of over $2 million spread across three tracks. The ARC-AGI-3 track offers $850,000 in total prizes, headlined by a $700,000 Grand Prize for the first eligible agent that achieves 100% on the fully private evaluation set. If no one claims it this year, the funds roll over to 2027.

Who is Tufa Labs, and how do they keep winning?

Tufa Labs is a highly specialized AI research lab completely dedicated to solving general intelligence benchmarks, specifically the Abstraction and Reasoning Corpus. They are the creators of the StochasticGoose framework. They previously placed 3rd in the ARC Prize 2025 global competition and have been a consistent presence at the top of every ARC leaderboard.

Their core approach leans into reinforcement learning rather than large language models. Their StochasticGoose system uses a CNN with reinforcement learning to predict which actions would cause frame changes, encoding 64x64 frames through a four-layer convolutional network. It achieved 12.58% and completed 18 levels during the preview competition. The milestone-winning "duck harness" submission is an evolution of this same philosophy.

The system uses a simple reinforcement learning approach to predict which actions will cause frame changes, enabling more efficient exploration than random selection. Lead Developer is Dries Smit, with Jack Cole as Adviser/Reviewer. ML Street Talk's Tim Scarfe traveled to Zurich to sit down with the Tufa Labs ARC-AGI-3 team , founder Benjamin Crouzier, with Jeroen Cottaar, Dries Smit, Stefano Viel, and Michal Tesnar , to work out what their leaderboard-topping system does and what the benchmark is really testing.

Why RL beats LLMs here

The results expose a structural mismatch. The top-scoring agents during the preview period were not language models. A system called StochasticGoose from Tufa Labs reached 12.58% using a reinforcement learning approach with convolutional neural networks. Meanwhile, frontier models at release scored: Anthropic Opus 4.6 at 0.50%, Google Gemini 3.1 Pro Preview at 0.40%, OpenAI GPT 5.4 at 0.20%, and xAI Grok-4.20 at 0.10%.

ARC-AGI-1 resisted pretraining scaling because base LLMs are limited to memorization and interpolative retrieval of patterns found in training data and cannot generalize to never-seen-before tasks. The latest generation of base LLMs still perform poorly on ARC-AGI-1. Test-time reasoning was the key innovation that enabled LLM-based systems to begin exhibiting non-zero fluid intelligence. ARC-AGI-3 now demands something beyond even that: an agent that learns the rules of a world it has never seen, from scratch, in real time.

The benchmark also separates two kinds of approaches. ARC-AGI-3 maintains two separate leaderboards with different rules. The Official leaderboard evaluates frontier models using general-purpose API systems not specially prepared for ARC-AGI-3. The Community leaderboard allows self-reported results and permits harnesses. This distinction matters because hand-crafted harnesses can dramatically inflate scores on known environments while failing completely on unseen ones. The Official leaderboard is designed to measure actual adaptability, not the developer's scaffolding. The Kaggle milestone prizes reward the community leaderboard track.

What the open-source release means

The primary mission of ARC Prize is to accelerate progress toward open AGI by making cutting-edge solutions freely available to the entire research community. They provide substantial prize money, dedicated infrastructure, and public recognition to incentivize participants who share reproducible methods openly.

With three notebooks now public, the research community gets a concrete starting point. The key ideas to study from the winning submissions:

• Action-change prediction , using a CNN to learn which clicks actually change the game state, rather than random exploration
• Sparse reward RL , handling environments where the only signal is a single success per level
• Harness design , how to wrap an environment interaction loop that doesn't blow up context or action budgets

The ARC Prize Foundation expects that the best ideas originating from harness research, if they are sufficiently general, will end up flowing behind the model API layer. For example, the original chain-of-thought research started out as a third-party harness from DeepMind wrapping GPT-3. The same pattern may be playing out here.

What comes next

The competition runs through November 2, 2026, with results announced December 4. A second milestone checkpoint on September 30 offers another $37,500 in prizes. The 85% grand prize threshold remained unclaimed across both prior competition years , and with the best current agent sitting at just over 1%, the $700K grand prize for ARC-AGI-3 looks safe for now.

The real story is what these open-source notebooks unlock. Every team building agents for the September milestone now has a working baseline. The question is whether the community can push from 1% toward something that actually looks like learning , and whether that path requires a fundamentally new architecture, or just more of what Tufa Labs has already figured out.