A small startup just achieved something that would have seemed impossible during the height of Silicon Valley's "bigger is better" AI era: matching a major lab's reasoning performance while cutting costs by orders of magnitude.
rekursiv.ai, emerging from Y Combinator's Summer 2026 batch, reached 59.8% accuracy on the notoriously difficult ARC-AGI-1 reasoning benchmark at a cost of $0.000035 per task in July. That performance essentially ties OpenAI's o4-mini model, which scored 58.7%, but at roughly 11,600 times lower cost, according to a company blog post published July 16, 2026. According to rekursiv.ai, the work came from what the startup describes as an autonomous AI research team that ran 235 experiments, testing various architectures and approaches without humans generating hypotheses along the way.
The result arrives at an inflection point for artificial intelligence development. Frontier laboratories spent much of 2024 and early 2025 pursuing accuracy by throwing massive amounts of compute at reasoning problems. OpenAI's o3 model approached human-level scores on ARC that December, but at implied costs running into thousands of dollars per task, TechCrunch reported at the time. Now a loose coalition of startups and independent researchers is racing toward something different: radical efficiency on those same benchmarks.
The Benchmark That Humbled Giants
François Chollet created ARC-AGI to test abstraction and reasoning through visual pattern puzzles. Humans solve these trivially. The largest language models struggled with them until recently. The benchmark suite now spans three versions. ARC-AGI-1 and ARC-AGI-2 measure static pattern inference from input-output examples. ARC-AGI-3, which launched in March 2026, introduced interactive environments testing exploration, goal inference, and planning efficiency, according to a technical report the ARC Prize Foundation published the following month.
The benchmark has become something of an industry standard for reasoning. Stanford's AI Index 2026, released in April, lists ARC-AGI among core reasoning metrics. Frontier labs including OpenAI, Anthropic, Google DeepMind, and xAI now report ARC performance in their model documentation.
Yet the difficulty remains severe. Humans solve 100% of ARC-AGI-3's public environments. Frontier AI systems scored below 1% as of March 2026, according to the foundation's technical report.
"Intelligence is the efficiency with which you're going to make sense of new things, of new tasks that you've never seen before," Chollet said in a March 25, 2026, Fast Company interview. That framing drove ARC Prize 2026 to restructure its $2 million competition around efficiency as much as raw accuracy.
The prize's verified testing policy, updated in 2026, caps each submission at a single run and $10,000 in compute costs. Teams must optimize per-task expense at retail API pricing, according to the policy documentation accessed in August 2026.
Mike Knoop, ARC Prize co-founder, put the stakes plainly to Fast Company: "ARC is literally the most important unbeaten benchmark in the world because it is the only really clear evidence that contradicts the scaling story that was so dogmatic in the Bay Area in 2023 and 2024."
Even OpenAI researcher Noam Brown acknowledged the gap in the same piece. "There are still important ways in which AI falls short of human intelligence," he said. "One of the clearest is the ability to adapt efficiently in novel settings, which ARC-AGI-3 is designed to test."
Multiple Paths to Efficiency
rekursiv.ai's approach centers on what the company calls autonomous AI scientists: systems that ideate, experiment, learn, and iterate without human oversight for each hypothesis. The team achieved multiple optimal points on the accuracy-versus-cost curve. They hit 71.4% accuracy at $0.018 per task, and 75.5% at $0.058 per task. Both results came in 16 to 19 times cheaper than peers achieving similar accuracy, according to the company's blog.
Their methods included iterative campaigns, improved baselines built on prior work with Tiny Recursive Models, and ensembles. Notably, they avoided per-task test-time training, according to the July 16 post. rekursiv.ai also transferred improvements from ARC-AGI-1 to achieve 17.5% single-model accuracy on the harder ARC-AGI-2 benchmark. The startup is open-sourcing some components, including tools they call configgle and sagent.
Pathway, a research lab building architectures beyond the dominant Transformer model, hit a different efficiency milestone on August 11, 2026. Its BDH-CQ model uses just 150 million parameters and processes reasoning in a latent space rather than generating long chains of thought. That model scored 29.5% pass@2 accuracy on ARC-AGI-1 at $0.00070 per task, according to the lab's research blog. Pathway positioned this as an independent reproduction and argued that recurrent latent computation can undercut token-heavy reasoning on cost.
A third approach emerged in a July 7, 2026, arXiv paper titled "Cost-Effective Agent Harnesses." Researchers used an open-weight model, DeepSeek V3.2, with an Explorer-Definer pipeline and what they call a Reflective Orchestrator. They achieved 67.25% pass@2 accuracy at $0.62 per task, with no fine-tuning specific to ARC. The paper demonstrated that architectural orchestration over commodity models can reach mid-60s accuracy for under a dollar per task.
These results look strikingly different from the high-compute regime. When OpenAI's o3 model approached 87.5% accuracy on ARC-AGI-1 in high-compute mode in December 2024, the ARC Prize Foundation noted that "you could pay a human to solve ARC-AGI tasks for roughly $5 per task." The implication was clear: o3's costs ran into the thousands per task, according to a foundation blog post from December 20, 2024.
Other labs are exploring related efficiency strategies. A paper under review at NeurIPS 2026, apparently posted around May or June, described a knowledge-centric self-improvement framework shifting the locus of learning from individual agents to a shared, curated knowledge base. That approach reported 82% ARC-AGI-1 accuracy at $46 per task using a Claude 4.5 Haiku backbone. While higher in absolute cost than rekursiv.ai's runs, it showed lower token spend compared to agent-centric baselines like Darwin Gödel Machine and HyperAgents, according to the project page.
Enterprise Implications
The efficiency breakthroughs on ARC are arriving alongside enterprise momentum around agentic AI more broadly. Google Research published quantitative scaling principles for agent systems on January 28, 2026, finding that multi-agent coordination helps on parallel tasks but can hurt performance on sequential ones. That finding is directly relevant to ARC-AGI-3's sequential exploration challenges. Microsoft Research, in a February 26, 2026, blog post, outlined failure modes in long-horizon corporate agent simulations and mapped required capabilities including memory, planning, and learning.
Frontier labs face mounting pressure to deliver on efficiency alongside raw capability. The ARC Prize Foundation wrote in its April 2026 technical report that ARC-AGI-3 is designed to remain unsaturated. It will emphasize action-efficiency and out-of-distribution novelty to counter overfitting.
A cross-benchmark study from late 2025 reported inference cost improvements of roughly five to ten times per year across multiple benchmarks, though that analysis is now perhaps more dated than ideal.
What Comes Next
Anthropic's Institute for AI Policy published a mid-2026 analysis titled "When AI builds itself," outlining scenarios where recursive self-improvement enables "100-person companies [to] do the work of 10,000- or 100,000-person organizations" as agent automation compounds. The report warned that progress bottlenecks will shift to code review and governance, and called for coordination mechanisms if recursive self-improvement accelerates beyond safe speeds.
rekursiv.ai's autonomous scientist approach represents one path toward that future. The startup framed its system as an "ideate, experiment, learn, repeat engine" in its July blog post. The company declined to disclose funding details but lists its Y Combinator batch as Summer 2026 on the accelerator's website. The site, accessed in August, explicitly claims "up to 10,000× lower cost" versus frontier LLMs on ARC.
Pathway, Symbolica, Sakana AI, and other newer labs are pursuing parallel bets on self-improvement architectures. Sakana's AI Scientist program was reported by third-party evaluators sometime in 2025 to generate research papers at $6 to $15 per paper with roughly 3.5 hours of human oversight. Those figures illustrate economics shifting beneath traditional R&D, though they are now over a year old and focused on paper generation rather than benchmark performance.
ARC Prize's $2 million purse and open-sourcing requirements are designed to accelerate reproducible efficiency breakthroughs. The foundation announced milestone awards on June 30, 2026, for both ARC-AGI-2 and ARC-AGI-3 tracks.
Fast Company reported in March that ARC Prize organizers expect labs to pivot further toward agentic qualities: goal inference, exploration, efficient planning. Chollet told the magazine that ARC-AGI-3 "measures on-the-fly reasoning rather than memory recall" and places efficiency at the center of what counts as intelligence.
The race has shifted, then. It is no longer purely about reaching human-level reasoning. The new contest is reaching it at a cost structure that makes autonomous research and recursive improvement economically viable at scale. Whether the first movers in efficiency will hold their advantage, or whether frontier labs will eventually match both accuracy and cost, remains an open question.
