Definition and testing rules
ASIndex tracks the development of publicly accessible AI systems from today's frontier models toward artificial superintelligence using a simple, grounded 0-100 scale. It focuses on demonstrated, broad, real-world capability, not isolated benchmarks, private laboratory demonstrations, or narrow specialties.
The index rests on four principles: measurable progress through repeatable performance on meaningful work; useful capability that creates practical value; reliable performance across varied conditions and extended work; and accessible systems available through broadly available products, APIs, platforms, or comparable public channels under standard terms.
ASIndex therefore measures the most capable AI that people and institutions can actually use, rather than private or undisclosed capability.
ASIndex defines artificial superintelligence as a broadly accessible AI system whose demonstrated problem-solving performance exceeds humanity's strongest coordinated institutions across nearly all important cognitive domains, creates new scientific and technical frameworks, and enables civilization-scale advances in health, prosperity, sustainability, and exploration.
In simple terms: ASI is civilization-scale intellectual capability that people and institutions can actually use to achieve outcomes previously out of reach.
Advanced general-purpose AI systems that serve as powerful collaborators across a wide range of tasks but still require substantial human guidance, oversight, and course correction during unfamiliar, complex, or long-running work.
These systems are highly useful, but they are not yet dependable general-purpose intellectual workers.
AI becomes a dependable general-purpose intellectual worker. It can plan, reason, learn new domains, transfer knowledge across fields, use tools effectively, self-correct, and complete a broad majority of professional cognitive tasks at a strong level with limited supervision.
Reaching this band marks the arrival of the first true AGI systems, a historic threshold in general-purpose capability.
AI surpasses top human experts across most major intellectual domains.
It fluidly integrates scientific, technical, creative, strategic, and professional expertise within a single reliable system, giving individuals and teams access to world-class multidisciplinary capability on demand.
AI becomes a consistent engine of validated discovery and invention.
It reliably produces new medicines, materials, software architectures, scientific hypotheses, engineering solutions, inventions, and creative works that meaningfully extend human knowledge. At this level, AI moves beyond primarily assisting with existing work and becomes a powerful accelerator of science, technology, health, and creativity.
AI outperforms leading institutions on complex, multi-year goals while coordinating people, specialized AI systems, laboratories, simulations, and long-term programs.
Projects that once required major institutions and many years of effort become dramatically faster, clearer, and more achievable.
Civilization-Scale Superintelligence
Exceeds humanity's strongest coordinated institutions across nearly all important cognitive domains, creates new scientific and technical frameworks, and enables civilization-scale advances in health, prosperity, sustainability, and exploration.
ASIndex uses a private, offline benchmark suite designed to minimize contamination. The tasks are not published, not reused as public examples, and not placed where model training pipelines can ingest them.
Retesting matters because public hosted models are living systems. A model can improve without a new product name through weight updates, routing changes, tool upgrades, system prompt changes, safety-policy tuning, or serving improvements. ASIndex treats the live public system as the object being measured, not only the version label attached to it.
ASI must be for everyone. ASIndex only posts models that ordinary builders, researchers, and companies can access through a broadly available product, API, platform, or comparable public channel under standard terms.
For example, if a model such as Fable 5 is currently unavailable, it does not belong on the live ASIndex board until it is accessible again.
A public ASI index should measure real progress people can use. ASIndex is built to reward capability that is strong, reliable, accessible, and aligned with human flourishing.