Poetiq, a Mountain View, California, startup that builds software on top of large language models, raised a $45.8m seed round, SiliconANGLE and Built In reported at the end of January 2026. FYRFLY Venture Partners and Surface Ventures led the round.
Y Combinator, 468 Capital, Operator Collective, Hico Ventures and Neuron Venture Partners also took part. No valuation was disclosed.
“For ARC-AGI 1 and 2, we used recursive self-improvement to produce specialized agents in a matter of hours.”Shumeet Baluja, co-founder and co-CEO, Poetiq
What Poetiq does
Poetiq describes what it builds as a meta-system: a layer that sits above existing frontier models rather than inside them. SiliconANGLE said it works with models such as Google's Gemini series and GPT-5.2 and with open-source models. Built In listed ChatGPT, Claude and Gemini. The company's website says the approach needs far less data than reinforcement learning and works with the models customers already use. It also mentions an LLM router that picks the best model for each query.
According to the company, a customer supplies a few hundred examples of a task rather than thousands or millions. The meta-system then generates and refines a specialised agent: the agent produces an answer, collects feedback, and improves the answer using the underlying model.
SiliconANGLE reported that the software also tries to cut inference costs. It determines when a model has gathered enough information to answer and stops further research queries, which avoids unnecessary computation.
Benchmark claims
Poetiq says its meta-system enabled GPT-5.2 to beat the previous best score on ARC-AGI-2 by 16%. ARC-AGI-2 is a benchmark of visual reasoning puzzles; SiliconANGLE said it has 1,000 of them, while Built In said only that the company recently reported results without giving scores. The figure is the company's claim and the sources do not say it was independently verified.
Shumeet Baluja, co-chief executive, said in the announcement: "For ARC-AGI 1 and 2, we used recursive self-improvement to produce specialized agents in a matter of hours." Built In also quoted him saying that large language models are impressive databases that encode a vast amount of humanity's collective knowledge, but are simply not the best tools for deep reasoning.
The founders
SiliconANGLE said the company was launched in 2025 by former Google DeepMind researchers Shumeet Baluja and Ian Fischer. Baluja established the unit's computer vision group, and Fischer helped develop technologies underlying its language models. The team includes other former Google researchers.
Poetiq's website lists a tagline of the fastest path to safe super intelligence. Another report, from The SaaS News, said the founders started the company in June 2025 and that the seed money will go to developing the meta-system, expanding integrations with frontier models and growing the team. That report described the product as designed to solve complex real-world problems that frontier models struggle with, and said it works with models including GPT, Claude, Gemini and LLaMA. The sources do not state the size of the team or whether Poetiq raised anything earlier.
Company website: poetiq.ai