Poetiq Secures $45.8M Seed Funding for AI LLM Optimization System

Poetiq did not come out of stealth quietly. It walked onto the floor in January 2026, Mountain View address on the label, $45.8 million in Seed capital on the table, and a very specific message humming under the noise. The models are not broken. The way we ask them to think is. Founded in June 2025 by Shumeet Baluja, PhD and Ian Fischer, both alumni of Google DeepMind, Poetiq was built on a simple refusal to accept that reasoning should cost weeks of retraining and millions in compute just to get a straight answer.

The story starts at DeepMind, where Baluja and Fischer watched frontier models crush pattern recognition and then stall out when problems required patience, structure, and second thoughts. Pre-training was the hammer. Everything looked like a nail. Poetiq took a different route. Build a meta system that sits above the models, orchestrates them, critiques them, pushes them back into the ring, and does it again until the answer holds up under pressure. No retraining. No allegiance. Just better thinking, cheaper.

That thinking turned into benchmarks. In December 2025, roughly six months after founding, Poetiq hit 54 percent accuracy on the ARC AGI 2 semi-private evaluation using Gemini 3 Pro at about $30 per problem. That alone edged past the field. Then GPT 5.2 landed. Poetiq integrated it within hours and pushed the public ARC AGI 2 score to 75 percent at under $8 per problem, clearing the human baseline and widening the cost gap in the same move. Reasoning got faster. The bill got lighter. That is not a metaphor. That is math.

Investors noticed. FYRFLY Venture Partners and Surface Ventures co-led the Seed, joined by Y Combinator from its S25 batch, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners. The round brought total funding to $45.8 million, no prior raises, no valuation disclosed, and a clear mandate to productize what was already working. This is Poetiq’s first public chapter, not a victory lap.

The team is still small enough to name. Alongside Baluja and Fischer are David Marwood as Founder and Head of Engineering, plus founding research scientists Saurabh Singh, Yair Alon, Michael Hale, and Ashwin Baluja. Seven people, decades inside Google and DeepMind between them, now focused on one uncomfortable truth. Enterprises have poured $30 to $40 billion into GenAI and most of it has not paid them back.

Poetiq is not trying to be the smartest model in the room. It is trying to make every model think twice. The recursive loop is running. The cost curve is bending. The next question is not whether this works. It is where it lands first.