Poetiq

Poetiq is an American artificial intelligence startup founded in June 2025 by Shumeet Baluja and Ian Fischer, both former AI researchers at Google DeepMind. The company develops a meta-system that orchestrates existing foundation models including ChatGPT, Claude, Gemini, and Llama, generating specialized agents for individual problems and recursively improving them for accuracy and cost efficiency. As of May 2026, Poetiq has raised $45.8 million in seed funding and has reported state-of-the-art results on the ARC-AGI-2 benchmark.

At a glance

• Founded: June 2025 in the United States by Shumeet Baluja and Ian Fischer. Y Combinator participant.
• Status: Private. Seed round closed early 2026.
• Funding: $45.8 million seed round, co-led by FYRFLY Venture Partners and Surface Ventures. Y Combinator, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners participated.
• CEO: Co-CEO structure. Shumeet Baluja and Ian Fischer share the chief executive role.
• Other notable leadership: Shumeet Baluja, co-CEO. Former CTO of Jamdat Mobile (IPO 2004) and a 21-year veteran of Google DeepMind, where he founded the mobile practice and started the fundamental computer-vision research group. Ian Fischer, co-CEO. Joined Google DeepMind through the 2015 acquisition of Apportable, where he was co-founder and CTO. Founding researchers and engineers include Yair Alon, Saurabh Singh, Michael Hale, Ashwin Baluja, and Michele Covell.
• Open weights: Not the company's principal output. Poetiq's product is a software meta-system layered on top of existing foundation models.
• Flagship outputs: The Poetiq meta-system, with reported state-of-the-art results on ARC-AGI-2 in late 2025 and early 2026.

Origins

Poetiq was founded in June 2025 by Shumeet Baluja and Ian Fischer, both alumni of Google DeepMind. Baluja's career predated DeepMind: he was Chief Technology Officer of Jamdat Mobile (which had its IPO in 2004) and subsequently spent 21 years at Google and DeepMind, where he founded the mobile practice and started the fundamental computer-vision research group. Baluja is a co-author on 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 the 2015 acquisition of Apportable, where he was co-founder and CTO. Apportable was a platform for porting iOS games to Android, an unusual on-ramp into the AI research career that produced his subsequent contributions at DeepMind on agentic and language-model research.

The founding thesis is that existing foundation models (ChatGPT, Claude, Gemini, Llama) are powerful capability primitives but are not optimally configured for specific problems. A "meta-system" that takes a problem description plus a few hundred examples and recursively constructs and improves agents on top of those foundation models can outperform any single model on the same task, at lower cost. The thesis explicitly contrasts with approaches that assume frontier capability requires ever-larger model training runs.

The company's first major external milestone was a state-of-the-art result on ARC-AGI-2, the abstract-reasoning benchmark created by François Chollet and now run through the ARC Prize Foundation. In December 2025, Poetiq established a SOTA on the ARC-AGI-2 semi-private evaluation set, topping Gemini 3 Deep Think (the previous leader) at half the cost per task. Following the December 2025 release of OpenAI's GPT-5.2, Poetiq incorporated the new model into its meta-system and reported a 75-percent accuracy on the public evaluation set, a 16-percentage-point improvement on the prior state-of-the-art.

The $45.8 million seed round closed in late January 2026, co-led by FYRFLY Venture Partners and Surface Ventures with participation from Y Combinator, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners. Y Combinator's Garry Tan publicly noted Fischer's prior YC W11 alumnus status from his Apportable era and welcomed Baluja into the YC family. The implied seed-round valuation has not been broadly publicly disclosed.

The team at the time of seed closure was reported as approximately six people, with the ARC-AGI-2 results characterized in industry coverage as a "lean squad outsmarts AI giants" headline.

Mission and strategy

Poetiq's stated mission is to build AI that "actually reasons" through orchestration and recursive improvement of agents on top of existing foundation models. The strategic premise reflects a structural distinction between foundation models (which provide raw capability) and the application-specific configuration of those models for individual problems (where Poetiq's meta-system is positioned).

The strategy combines three threads. The first is the meta-system itself, which sits as a layer above multiple foundation models and constructs problem-specific agents. The second is recursive self-improvement of agents within the meta-system, where the system iteratively improves its own approach to a given problem. The third is the deliberate model-neutral positioning, which keeps Poetiq's product compatible with any foundation model rather than dependent on any single provider.

The competitive premise is that capability gains at the meta-system level can offset gaps at the foundation-model level, that frontier benchmark performance is therefore not exclusively the property of the labs training the largest models, and that customers will pay for application-specific agent construction tooling. The premise depends on Poetiq's meta-system maintaining its capability lead as foundation-model providers improve their own first-party agentic offerings.

Models and products

• Poetiq meta-system. Software platform that takes a problem description plus a small number of examples and constructs problem-specific agents on top of foundation models including ChatGPT, Claude, Gemini, and Llama. Supports recursive improvement of agents within the meta-system.
• ARC-AGI-2 results. Used as the principal external benchmark for capability claims. Poetiq established state-of-the-art results in December 2025 and improved them following the GPT-5.2 release.
• Customer-facing access. Specific commercial product details, pricing, and customer disclosures have not been broadly publicly disclosed as of May 2026.

Distribution channels appear to be direct enterprise sales, with the principal marketing vector being the ARC-AGI-2 benchmark results and the broader research-community visibility they generate. Specific go-to-market details are not public.

Benchmarks and standing

ARC-AGI-2 is the principal benchmark associated with Poetiq's standing. The December 2025 SOTA on the semi-private evaluation set established the company's research credibility. The post-GPT-5.2 result of 75 percent on the public evaluation set widened the lead.

Standing within the broader frontier-AI research community draws on the ARC-AGI-2 results, the founder team's Google DeepMind credentials, and the lean-team headline. Industry coverage has characterized Poetiq as a research-credibility-anchored insurgent that demonstrated meaningful frontier-AI capability with substantially smaller capital and team than the foundation-model labs.

For commercial standing, the data is sparse. Whether the meta-system approach converts to durable commercial advantage depends on customer acquisition, retention, and pricing power against the in-house agentic offerings from OpenAI, Anthropic, and Google DeepMind.

Leadership

As of May 2026, Poetiq's senior leadership includes:

• Shumeet Baluja, co-Chief Executive Officer and co-founder. Former CTO of Jamdat Mobile and 21-year Google DeepMind veteran. Founder of DeepMind's mobile practice.
• Ian Fischer, co-Chief Executive Officer and co-founder. Joined Google DeepMind via the 2015 acquisition of Apportable, where he was co-founder and CTO.
• Yair Alon, founding research scientist.
• Saurabh Singh, founding research scientist.
• Michele Covell, research scientist.
• Michael Hale, founding research engineer.
• Ashwin Baluja, founding research engineer.

The team at seed closure was reported as approximately six people. Subsequent hiring has not been publicly broken out in detail.

Funding and backers

Poetiq's funding history through May 2026 consists of one disclosed round:

• Seed (January 2026): $45.8 million, co-led by FYRFLY Venture Partners and Surface Ventures. Y Combinator, 468 Capital, Operator Collective, Hico Ventures, and Neuron Venture Partners participated.

The seed-round size is large for a six-person team but consistent with the meta-system thesis and the ARC-AGI-2 capability demonstration. The investor base spans early-stage US VCs (FYRFLY, Surface, 468) and operator-network investors (Operator Collective).

The implied seed-round valuation has not been broadly publicly disclosed in primary coverage. Industry estimates have placed Poetiq below the $500 million range typical of comparably-funded foundation-model insurgent peers.

Industry position

Poetiq occupies a distinctive position as a research-credibility-anchored agentic AI startup with state-of-the-art ARC-AGI-2 results and a deliberately model-neutral architecture. The combination of the founder team's Google DeepMind tenure, the lean-team capability demonstration, and the recursive-improvement framing produces a profile that the foundation-model labs do not directly match.

Industry coverage has characterized the company as one of the principal insurgents in the agentic-AI category, with the ARC-AGI-2 results providing strong external evidence of capability. The principal strategic-execution risks identified are the dependence on third-party foundation models (which constrains pricing power and exposes the company to API-cost shifts), the competitive pressure from in-house agentic offerings at OpenAI, Anthropic, and Google DeepMind, and the conversion of research-benchmark performance into durable commercial product positioning.

Competitive landscape

Poetiq competes for agentic AI workload across several categories:

• Foundation-model labs. OpenAI, Anthropic, Google DeepMind, Meta AI / FAIR, and xAI all provide first-party agentic and reasoning capabilities. Poetiq's meta-system layers above them.
• Ndea. The closest peer on ARC-AGI-research positioning, with founders François Chollet and Mike Knoop directly affiliated with the ARC-AGI benchmark. Ndea pursues program synthesis as a core methodology rather than meta-system orchestration.
• Other agentic-AI insurgents. H Company, Cognition AI, Reflection AI, and other agentic startups compete for enterprise and developer customers.
• Coding agents. Cursor (Anysphere), Codeium / Windsurf, and Magic target a more specialized vertical but compete for similar customer attention on technical workloads.
• Cloud-provider agentic platforms. AWS, Google Cloud, and Microsoft Azure compete on enterprise agent orchestration through their respective managed-AI services.

Outlook

Several open questions affect Poetiq's trajectory in 2026 and 2027:

• Continued ARC-AGI-2 progression as foundation models continue to improve. Whether the meta-system advantage holds against larger frontier model releases is the principal capability question.
• Commercial customer traction, with named enterprise customers and revenue trajectory among the watchable signals.
• Subsequent fundraising; the seed-round structure suggests a Series A at a higher valuation if commercial traction holds.
• The competitive dynamic with Ndea on the ARC-AGI-positioned research narrative, particularly as both companies publish further work.
• The depth of integration with each supported foundation-model provider, given the dependency on third-party APIs.
• The team's hiring trajectory and whether the lean-team positioning persists as the company scales commercially.

Sources

• SiliconANGLE: Poetiq nabs $45.8M in seed funding for its LLM-enhancing meta-system. Seed funding announcement.
• PR Newswire: Poetiq Raises $45.8M for AI Meta-System, Surpasses Top LLMs on Industry Benchmark. Funding and benchmark coverage.
• Puck: How Poetiq's Six-Person Team Beat Google at A.I.. Lean-team and capability framing.
• The AI Insider: Poetiq Raises $45.8M for AI Meta-System. Industry coverage.
• Operator Collective: The Meta-System Making AI Actually Reason. Investor perspective.
• Poetiq official site. Company reference.
• ARC Prize Foundation. Benchmark organization.