Poetiq, a Mountain View, CA-based developer of an AI meta-system for LLM optimisation, has raised $45.8 million in a financing series led by FYRFLY Venture Partners and Surface Ventures.
This round saw also participation from Y Combinator, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners.
The funding comes after Poetiq’s strong performance on ARC-AGI-2, a key industry benchmark for machine reasoning and progress toward artificial general intelligence (AGI).
Read More:PranaX Raises $17M in Series A Funding
The company plans to use the funds to grow its operations and advance its development work.
Poetiq develops an AI meta-system that enhances the performance of large language models (LLMs). It works with any leading LLM—like OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, or Meta’s Llama—helping them learn faster and tackle more complex problems. Clients provide a problem, and a few hundred examples, rather than thousands or millions, and Poetiq’s system creates an agent specialized in that problem, improving it over time to become more accurate and cost-efficient.
The Abstraction and Reasoning Corpus (ARC-AGI), created in 2019 by AI researcher François Chollet, tests an AI’s ability to generalize problem-solving like a human. In early December, Poetiq set a new state-of-the-art (SOTA) on the ARC-AGI-2 semi-private evaluation, outperforming Gemini 3 Deep Think at half the cost per task using their system on Gemini 3 Pro.
Shortly after, OpenAI released GPT-5.2, which Poetiq quickly integrated into their system, achieving 75% accuracy on the public evaluation—a 16-point improvement over the previous SOTA. OpenAI co-founder Greg Brockman tweeted that Poetiq is “exceeding the human baseline on ARC-AGI-2 with GPT-5.2.
“LLMs store a huge amount of human knowledge,” said Shumeet Baluja, co-CEO of Poetiq.
“But they aren’t ideal for deep reasoning, which makes improving their problem-solving slow and costly. For ARC-AGI 1 and 2, we used recursive self-improvement to create specialized agents in just hours, showing how we can tackle problems that LLMs alone find too hard or expensive.”
A MIT study of 300 public AI projects, published in August 2025, highlights the need for Poetiq. Even though companies have spent $30–$40 billion on generative AI, researchers found that 95% of organizations see “zero return.” Projects that struggle to deliver ROI are ideal for Poetiq, which can enhance the reasoning abilities of any LLM, including in-house or proprietary models.
“Remarkably, Poetiq reached the top of ARC-AGI just six months after launching,” said Philipp Stauffer, General Partner at FYRFLY Venture Partners. “Instead of competing with leading models, their six-person team found a way to get more intelligence from every LLM.
Poetiq will be essential for companies looking to make AI work in real-world business applications.”
“Poetiq is one of the few AI startups that doesn’t have to compete with leading models or choose sides,” said Gyan Kapur, co-Managing Partner at Surface Ventures. “It can improve any LLM, any AI platform, and any AI use case, delivering better performance at lower cost by building on top of foundation models—a truly unique position.”
About Portiq
Poetiq was founded in June 2025 by co-CEOs Shumeet Baluja, PhD, and Ian Fischer, both former AI researchers at Google DeepMind. The company is advancing toward AGI with a meta-system that makes leading LLMs more innovative and more cost-effective for real-world problems. Using recursive self-improvement, Poetiq creates AI agents that solve tasks too difficult or expensive for LLMs alone, and these agents improve with each problem they tackle.
Baluja was previously CTO of Jamdat Mobile (IPO 2004) and spent 21 years at Google DeepMind, founding their mobile practice and starting their computer vision research group, with contributions to over 170 patents. Fischer joined DeepMind through its 2015 acquisition of Apportable, where he was co-founder and CTO.
Read More:Oracle set to raise up to $50 billion in debt and equity this year
