Poetiq has closed a $45.8 million Seed funding round co-led by FYRFLY Venture Partners and Surface Ventures. The round included participation from Y Combinator, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners. The AI startup, founded by former Google DeepMind scientists, emerged from stealth mode after achieving state-of-the-art (SOTA) results on ARC-AGI-2, an industry benchmark for machine reasoning and progress toward artificial general intelligence (AGI).
The funding comes as Poetiq demonstrates its ability to help large language models solve more complex problems faster. The company's meta-system works alongside any frontier LLM, including OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and Meta's Llama, to improve their reasoning capabilities.
The timing of Poetiq's funding aligns with growing concerns about enterprise AI returns. An MIT study of 300 public AI implementations, published in August 2025, found that despite $30 to $40 billion in GenAI investments, 95% of organizations are "getting zero return." Poetiq targets these underperforming use cases by improving the reasoning capabilities of any LLM, including proprietary, in-house models.
Current solutions for improving LLM performance require pre-training and post-training through reinforcement learning (RL), a process that takes weeks and remains prohibitively expensive for most companies. Poetiq offers an alternative approach: clients provide a problem and a few hundred examples instead of the thousands or millions required for fine-tuning or RL post-training. The company's meta-system then generates a specialized agent that recursively improves to become more accurate and cost-efficient.
Poetiq will use the funding to scale its technology and expand its reach to enterprises struggling with AI implementation. The company's approach allows it to work across diverse use cases by sitting on top of foundation models, providing better performance at lower costs.
The meta-system creates agents that specialize in solving specific problems and improve with each iteration. This recursive self-improvement process can produce specialized agents in hours rather than weeks.
Shumeet Baluja, PhD, and Ian Fischer co-founded Poetiq in June 2025 as co-CEOs. Both are former AI researchers at Google DeepMind who identified fundamental limitations in how frontier LLMs handle complex reasoning tasks. Baluja previously served as CTO of Jamdat Mobile (IPO 2004) and spent 21 years at Google DeepMind, where he founded their mobile practice and started their fundamental computer vision research group. He has contributed to more than 170 patents in neural networks, machine learning, and applications, and is one of the originators of YouTube's copyright system.
Fischer joined Google DeepMind through its 2015 acquisition of Apportable, a platform that ported iOS games to Android, where he was co-founder and CTO. Working together at DeepMind, the pair noticed that frontier LLMs struggled to solve most hard problems, and even many easy ones.
"LLMs are impressive databases that encode a vast amount of humanity's collective knowledge," said Shumeet Baluja, co-CEO of Poetiq. "They are simply not the best tools for deep reasoning. That's why efforts to improve their problem-solving skills are so slow and expensive. For ARC-AGI 1 and 2, we used recursive self-improvement to produce specialized agents in a matter of hours. It demonstrates how much we can help with problems that have been too hard or too expensive for LLMs alone."
The funding round reflects investor confidence in Poetiq's approach to improving AI reasoning. The company topped the ARC-AGI-2 benchmark, which measures an AI's "human-like generalization" of problem-solving skills. The Abstraction and Reasoning Corpus, developed in 2019 by AI researcher François Chollet, has become a standard for evaluating progress toward AGI.
In early December, Poetiq established a SOTA on the ARC-AGI-2 semi-private evaluation set, surpassing Gemini 3 Deep Think at half the cost per task. The company achieved this using its system on top of Gemini 3 Pro. Within days of OpenAI releasing GPT-5.2, Poetiq incorporated the model and achieved a new SOTA at 75% accuracy on the public evaluation set, a 16 percentage point improvement. OpenAI co-founder and President Greg Brockman responded to this achievement on social media, noting that Poetiq is "exceeding the human baseline on ARC-AGI-2 with gpt-5.2."
"That Poetiq managed to top ARC-AGI within six months of launching is remarkable," said Philipp Stauffer, General Partner at FYRFLY Venture Partners. "Rather than compete against frontier models, their team of six found a way to coax more intelligence from every LLM available. Poetiq will be a must-have for companies trying to make AI work for real-world business applications."
"Poetiq is one of the rare AI startups that doesn't need to outcompete frontier models or pick sides," added Gyan Kapur, co-Managing Partner at Surface Ventures. "It can enhance any combination of LLMs, any native AI platform, and any AI use case. Poetiq can provide better performance at lower costs across diverse use cases by sitting on top of foundation models, and that is a unique position to be in."
