How a Tiny AI Startup Outshone Google and Its Gemini 3

In a technological landscape dominated by giants like Google, a small AI startup proves that size isn’t everything. Poetiq, a team of only six members, has just surprised the industry by surpassing Google’s Gemini 3 model on a particularly demanding reasoning test. This unprecedented achievement raises many questions about the future of artificial intelligence models and the strategies needed to remain competitive.

While Gemini 3, developed by Google DeepMind, is presented as a major breakthrough in artificial intelligence with its multimodal capabilities and advanced reasoning, Poetiq demonstrates that an innovative approach, less focused on raw power, can trigger a true disruption. Their system relies on intelligent and iterative orchestration of existing models, including Gemini 3 Pro itself, to significantly improve the quality of responses in complex tasks.

This dynamic reflects an important evolution in AI competition where the efficiency of methods and the ability to quickly integrate new technologies become as crucial as the development of ever larger models. The startup Poetiq, through its transparency and agility, imposes a new paradigm in the race for technological innovation at the heart of artificial intelligence in 2025.

A small AI startup challenges Google: behind this exceptional performance

The AI startup Poetiq, despite its small size, has just achieved a remarkable feat. On the ARC-AGI-2 test, designed to challenge the logic and generalization of artificial intelligences, Poetiq scored an impressive 54%. This result clearly surpasses the 45% claimed by Google for its flagship Gemini 3 Deep Think model. This outperformance marks a key milestone, as it opens the way to a new understanding of AI performance in complex tasks.

This test does not merely gauge simple calculation or memorization capabilities. It evaluates more sophisticated skills:

• Recognition and manipulation of shapes
• The ability to establish analogies
• Abstract reasoning and formal logic
• Generalization beyond training data

The fact that Poetiq beats Google with a model based on the same foundation proves that artificial intelligence no longer depends solely on the model’s power, but also on how its use is orchestrated and optimized. This performance calls for deep reflection on development strategies in AI competition, highlighting the richness of alternative methods beyond just increasing parameters.

An innovative methodology: orchestrating models rather than simply improving them

Unlike traditional approaches that aim to create ever larger and more complex models, Poetiq has bet on an innovative metasystem. This system does not rely solely on an improved version of Gemini 3 Pro, but intelligently orchestrates multiple existing models, including Gemini 3, to maximize results.

The operation of this metasystem is based on a four-step iterative loop:

1. Generation: initial production of the response from the models.
2. Critique: analysis and critical evaluation of the generated response.
3. Refinement: improvement of the response based on critical feedback.
4. Verification: final validation to ensure quality and coherence.

This method, although simple, allows the full potential of existing models to be exploited without resorting to heavy and costly training. It also offers the advantage of rapid adaptation, with the ability to integrate improvements from future models in just a few hours.

Aspect	Google Gemini 3 Approach	Poetiq Metasystem
Nature of the system	Powerful monolithic model	Orchestration of multiple models
Cost per task	About 77 dollars	About 30 dollars
Adaptation to novelties	Heavy and long retraining	Adaptation in a few hours
Transparency	Limited access to internal tools	Published and accessible code

The ARC-AGI-2 test: a real challenge for technological innovation in AI

The ARC-AGI-2 test, created by researcher François Chollet, is recognized as one of the most demanding for measuring advanced artificial intelligence capabilities. Rather than evaluating tasks suited to traditional machine learning, this test emphasizes generalization and the ability to resemble human intelligence.

Key points assessed by ARC-AGI-2 include:

• The ability to identify non-trivial patterns in visual and symbolic environments.
• Solving problems requiring deep analogies.
• Developing strategies from limited data.
• Logical reasoning in varied and abstract contexts.

These skills remain a major challenge for most artificial intelligences, even the most advanced. Historically, models often stagnated around very low scores, sometimes below 5%. The rapid progress seen with Poetiq, which now reaches 54% in such a short time, illustrates a significant leap and reveals the combined effects of methodical orchestration and emerging technology capable of providing new flexibility.

Why is this acceleration in AI performance on ARC-AGI-2 so significant?

Several factors explain this spectacular advance:

1. Refinement of prompts and interaction strategies: the way questions are posed and answers formulated can significantly influence result quality.
2. Integration of critical iteration loops: Poetiq’s critical and iterative improvement phase reduces errors and increases accuracy.
3. Modularity in model management: combining several models to leverage their individual strengths creates a synergistic effect.
4. Democratization of code: opening their solution has stimulated community research and development, accelerating progress.

Factor	Impact	Application in Poetiq
Prompt refinement	10-15% improvement in scores	Use of task-specific prompts
Critical iteration loop	8-12% gain in accuracy	Successive critique and improvement of answers
Model modularity	Increased performance synergy	Orchestration of Gemini 3 Pro and other models
Code democratization	Acceleration of innovations	Open source publication of solvers

Reduced costs and transparency disrupt the traditional hierarchy

Beyond the figures in terms of performance, Poetiq’s approach creates a new dynamic in the artificial intelligence market. Resource-efficient management is a crucial strategic component that redefines expected standards.

Here’s how Poetiq’s strategy offers a significant competitive advantage:

• Cost reduction: completing a task for 30 dollars instead of 77 represents a significant saving, especially for large-scale industrial deployments.
• Open access to source code: publishing algorithms promotes broader collaboration, academic research, and accelerates technological innovation in the sector.
• Flexibility of adaptation: the method can quickly integrate new advances in models, avoiding the usual burdens of retraining.

This transparency and cost control are significant in a sector where giants like Google often prefer to maintain exclusive control over their internal technologies. Poetiq’s ability to upset this traditional hierarchy reflects a profound change as AI competition intensifies worldwide.

The impact on the future of artificial intelligence and machine learning models

This breakthrough by the AI startup Poetiq raises major questions about future directions in artificial intelligence development. One thing is certain: advances no longer come solely from massive model extensions, but also from methodological and strategic innovations.

Lessons learned from this success can be applied to several fields:

• Automated planning: systems that break down complex tasks into more manageable subtasks could benefit from these orchestration methods.
• Software development and coding: the ability to iteratively refine and correct significantly improves the efficiency of AI programming assistants.
• Advanced information retrieval: adaptive metasystems can more precisely guide queries in multi-format databases.

By modernizing the use of models rather than seeking to create the largest, the startup illustrates a path toward broader democratization of artificial intelligence. This approach in turn attracts growing interest from giants, who are now closely monitoring this type of disruptive innovation.

Application domain	Expected benefits	Concrete examples
Planning	Optimization of complex processes	AI project management for logistics
Coding	Improvement of generated code quality	Intelligent programming assistants
Research	Increased relevance of responses	Hybrid text-image search systems

Transparency that attracts the scientific community and fosters collaborative innovation

The open publication of source code by Poetiq is a major contributor to its success and marks a turning point in how AI competition is perceived. This transparency encourages:

• Independent performance validation, facilitating trust and recognition.
• Participation of external researchers, promoting healthy and constructive competition.
• Rapid sharing of best practices and accelerated dissemination of technological innovation.

In contrast to giants who often maintain restricted access to their internal technologies, this approach offers an ethical and pragmatic alternative. A snowball effect is already observed with a multiplication of contributions on collaborative platforms and an overall enrichment of the machine learning field.

Gemini 3: a turning point in artificial intelligence despite growing competition

Google’s Gemini 3 model remains a must-have breakthrough in the sector. With its extended multimodal capabilities, it excels in simultaneous analysis of texts, images, videos, sounds, and even code. This makes it a versatile tool, suited to a multitude of complex applications.

Nevertheless, despite these advantages, Gemini 3 faces significant challenges, notably:

• The increasing complexity of its training and optimization.
• High operating costs, sometimes limiting accessibility.
• A difficulty in maintaining stable logic on certain types of abstract questions.

The fact that a modest startup like Poetiq manages to deploy more effective orchestration emphasizes that even the most advanced models must evolve toward better process integration and a more modular approach.

Advantages of Gemini 3	Limitations encountered
Advanced multimodal understanding	High operational costs
Advanced reasoning	Logic sometimes unstable on certain tests
Wide deployment via Google API	Restricted source code access

Improvement prospects for Google and the AI sector

To remain a leader in this race, Google will need to adapt its strategies. This notably involves:

• Flexible integration of external models.
• Enrichment of self-evaluation loops to ensure answer reliability.
• Greater openness to the scientific community.

If these changes are not adopted, more agile and innovative players risk eating into significant market shares in the near future.

Lessons from Poetiq’s success for global AI competition

Faced with the historic dominance of technological behemoths, the Poetiq scenario is a demonstration that disruption is possible thanks to targeted and bold technological innovation. Some key lessons can be drawn:

• Methodological creativity takes precedence over raw power: intelligently thought-out orchestration is better than just scaling up the model.
• Operational agility: the ability to quickly integrate progress is a vital strategic factor.
• Transparency strengthens trust and accelerates innovation, creating a virtuous circle.
• Democratization of AI tools: providing open access encourages new ideas and the emergence of unexpected competitors.

These lessons upset the traditional sector hierarchy and encourage a more open competition, the benefits of which are already reflected in the rise of new human-scale actors.

Lesson	Implication for the future	Example from Poetiq
Methodological creativity	Reduced dependence on monster models	Metasystem orchestration instead of a single model
Agility	Responsiveness to rapid changes	Adaptation in a few hours
Transparency	Acceleration of community innovations	Publication of source code
Democratization	Market opening to diverse players	Making tools available as open source

The transformation of uses and expectations regarding artificial intelligence technologies

The unexpected success of the startup Poetiq changes perceptions about what artificial intelligence can truly achieve in the context of 2025. Companies, researchers, and end users are revising their priorities:

• Demand for a more logical and coherent AI: tasks requiring deep understanding and reasoning are the new frontiers.
• Growing importance of modularity: systems must be easily adaptable to meet varied needs.
• Preference for transparent and accessible AI: the community now values openness and collaboration over industrial secrecy.

These changes outline a new era where artificial intelligence will no longer simply be a technological tool, but a flexible and reliable partner integrating the best advances in machine learning and emerging technologies.

What is the ARC-AGI-2 test?

The ARC-AGI-2 test is a benchmark designed to evaluate the ability of artificial intelligences to reproduce complex reasoning, including pattern recognition, abstract logic, and generalization.

How did Poetiq surpass Gemini 3 despite its small size?

Poetiq developed a metasystem that orchestrates multiple AI models, including Gemini 3 Pro, following an iterative loop of generation, critique, refinement, and verification, which significantly improves performance without costly training.

Why is transparency an advantage in AI competition?

Transparency allows validation of performance, invites scientific collaboration, and accelerates innovations, unlike restricted access which can limit technology evolution.

What are the main challenges faced by Gemini 3?

Gemini 3 excels in multimodality and reasoning but faces high costs, sometimes unstable logic, and difficulty opening its source code for broad collaboration.

What lessons can the AI industry draw from Poetiq’s example?

The industry should foster methodological creativity, agility, transparency, and democratization of tools to remain competitive amid emerging innovative and agile players.