Problem

Businesses need reliable reasoning systems for practical workflows, but the prevailing way to improve large language models—pretraining and reinforcement-learning post-training—is slow, expensive, and data-hungry. Poetiq says RL often requires millions of data points and is practical mainly in areas such as coding and mathematics, where synthetic data is cheap; most business domains lack that advantage. The resulting pain is particularly acute for organizations that need specialized reasoning but cannot afford weeks of training, massive compute, or proprietary datasets at scale.

The central use case is turning a business problem and a few hundred examples into a reliable, task-specific reasoning capability. Rather than asking companies to retrain a model, Poetiq targets hard, domain-specific workflows where existing LLMs have knowledge but lack dependable problem-solving procedures.

Product / Service

Poetiq is an AI meta-system that sits on top of existing LLMs rather than modifying their weights. A client supplies a problem and a few hundred examples; Poetiq generates a specialized agent and recursively improves it for accuracy and cost efficiency. Its architecture can automatically assemble the surrounding harness—code, prompts, tools, and search strategies—needed for a particular task.

The benefit is a model-agnostic reasoning layer that can work with the LLMs a customer already uses, including frontier and open-source models, while requiring far less data than fine-tuning or RL post-training. Public materials emphasize work with select enterprise partners and an early-access model rather than a fully described self-serve pricing or deployment plan.

Market

Poetiq competes in AI reasoning infrastructure and the broader category of AI platforms and business-productivity software. Its adjacent competitive context includes frontier-model providers such as OpenAI, Anthropic, Google, and Meta, whose models Poetiq can use, as well as conventional reinforcement-learning and fine-tuning stacks. The available materials do not identify a named, like-for-like direct competitor; Poetiq’s stated differentiation is a model-agnostic layer that improves the performance of multiple underlying models.

The company appears to be early but well funded and technically validated. It was founded in 2025 by former Google DeepMind researchers, raised $45.8 million in seed funding in January 2026, and reports state-of-the-art results on ARC-AGI-1 and ARC-AGI-2 as well as six other benchmarks, without fine-tuning model weights. PitchBook labels the company as generating revenue, although exact revenue and customer figures are not disclosed, so it should not be treated as a mature-scale commercial business yet.