General artificial intelligence — or AGI for Artificial General Intelligence — is on everyone's lips. Sam Altman (OpenAI) announced he knows how to build it. Dario Amodei (Anthropic) speaks of a "nation of geniuses in a datacenter" by 2027. Demis Hassabis (DeepMind) remains more cautious but doesn't rule anything out. Meanwhile, part of the public oscillates between fascination and concern.

So, where do we really stand? Is AGI just around the corner, or is it a cleverly maintained technological mirage? This article provides an update, without unnecessary jargon, on the state of research in 2026.

What exactly is AGI?

Let's start at the beginning. AGI — Artificial General Intelligence or Artificial General Intelligence — refers to a system capable of performing any human cognitive task, with a performance level at least equivalent to that of an average adult.

It's not just a smarter chatbot. It's an AI that could write an article, diagnose a disease, learn a foreign language in a few hours, or design a bridge — all without having been specifically trained for each of these tasks.

Today, the AIs we use (ChatGPT, Claude, Gemini) are narrow AIs : they excel in their specific domain but are unable to transfer their skills from one area to another. AGI, on the other hand, would be capable of generalizing — exactly like a human.

Where does research stand in 2026?

OpenAI: Leading the Race

OpenAI remains the most publicized lab. In January 2025, Sam Altman declared: "We now know how to build AGI." Since then, GPT-5 was launched (August 2025), demonstrating "expert-level" reasoning capabilities in mathematics, science, and programming. More recently, in February 2026, GPT-5.3-Codex reached a new milestone in autonomous coding.

In April 2026, OpenAI even published its five principles for AGI development, a sign that the company considers the topic concrete enough to warrant a public framework. According to prediction markets (Kalshi, April 2026), there would be 55% chance that OpenAI reaches AGI by 2030.

But beware: "knowing how to build AGI" doesn't mean it's already been built. Between the roadmap and the final product, there's a world of difference — and billions of dollars in compute.

Anthropic: The Rising Challenger

Founded by former OpenAI executives, Anthropic has taken a unique position in the race for AGI. Its CEO, Dario Amodei, made headlines at the 2026 Davos Forum by stating that a "nation of geniuses in a data center" was likely by 2027. A bold prediction, but not without basis.

In February 2026, Anthropic launched Claude Opus 4.6, a model that demonstrated unprecedented capabilities in autonomous coding and multi-step reasoning. Weeks later, Claude Sonnet 4.6 — their mid-range model — matched the performance of premium models at a quarter of the price. Their Claude Code tool was adopted internally by Microsoft, despite being a direct competitor with GitHub Copilot.

Anthropic's philosophy stands out for its obsession with safety (safety). The company structures its research around Constitutional AI, an approach that aims to align models with explicit ethical principles rather than solely human feedback.

DeepMind (Google): The Methodical Approach

Less flashy than its competitors in its communication, DeepMind is nonetheless a central player. Demis Hassabis, its CEO and co-founder, has always advocated for a scientific and cautious approach to AGI. For him, AGI is not a sprint but a marathon — and he prefers to speak of 'levels' rather than a finish line.

It was DeepMind, in an article published in November 2023, that proposed the most rigorous reference framework for measuring progress towards AGI. This framework is now indispensable.

DeepMind's 5 Levels of AGI

To move beyond the binary 'AGI or no AGI' debate, DeepMind published a foundational article in November 2023, proposing a 5-level framework. This framework, now adopted by a large part of the scientific community, allows for a more nuanced measurement of AI progress.

Level 1 — Emerging (Emerging AGI): AI equals or slightly surpasses an unskilled human across a wide range of tasks. This is the level achieved by current LLMs like ChatGPT, Claude, or Gemini. They are 'emerging' because they show sparks of generalization but remain inconsistent.

Level 2 — Competent (Competent AGI): AI reaches the 50th percentile of skilled adults on most cognitive tasks. This level has not yet been publicly achieved, but some researchers believe that 2025-2026 models are approaching it on a subset of tasks (mathematical reasoning, coding).

Level 3 — Expert (Expert AGI): AI ranks in the 90th percentile of skilled adults. At this stage, it could write scientific papers, make complex medical diagnoses, or design entire software systems.

Level 4 — Virtuoso (Virtuoso AGI): AI reaches the 99th percentile. It surpasses almost all humans in most cognitive domains. This is the level many imagine when they think of "fully realized" AGI.

Level 5 — Superhuman (Superhuman AGI): AI surpasses 100% of humans on all tasks, including those that humans cannot perform. This level is also known as ASI (Artificial Superintelligence). It remains purely theoretical.

What's striking is that we are still at Level 1. Even the most advanced models of 2026 — GPT-5.3, Claude Opus 4.6 — remain, according to this framework, "emergent" AIs. They excel in certain narrow tasks (Level 3 or 4 in coding, for example) but do not generalize enough to merit Level 2.

Hype vs. Reality: How to Make Sense of It?

With statements as ambitious as those from Sam Altman or Dario Amodei, it's easy to get carried away. But it's crucial to distinguish real progress from the hype.

What is real. Language models are progressing at an impressive rate. In three years, we've gone from GPT-3.5 (which struggled with simple reasoning) to systems capable of solving Olympiad-level problems, coding complete applications, and reasoning through dozens of steps. Autonomous agents — AIs capable of using a computer like a human — are a technical reality in 2026. Claude Code, Anthropic's coding tool, has been adopted internally by Microsoft, despite being a direct competitor to GitHub Copilot. This is significant. And solutions like Delos Workers are already showing concrete early signs: these autonomous AI agents execute complex tasks end-to-end — web search, data analysis, content generation — without constant human supervision. We're not at AGI yet, but we're seeing the first building blocks.

What's overhyped. The "imminent" AGI has been a recurring theme since 2023. The reality is that current systems remain fundamentally limited: they hallucinate, lack persistent memory, do not understand the physical world, and possess neither intention nor consciousness. As Yann LeCun, Meta's Chief AI Scientist, reminds us: "If you want to work on AGI, don't work on LLMs." Language models, however impressive, remain machines that predict the next word — not thinking entities.

The question of compute. Achieving AGI won't just be a matter of algorithms. The computational power requirements are colossal, and costs are skyrocketing. Between the roadmap and the final product, there's a world of difference — and billions of dollars in infrastructure.

Conclusion: Neither Myth Nor Miracle — A Path to Build

So, AGI: myth or reality?

Neither. AGI is not a myth — progress is tangible, and the direction is clear. But it's not an imminent reality either. We are at the beginning of a journey, not the end.

What's important to remember is that AGI won't arrive overnight. It will emerge gradually, in successive stages, as proposed by DeepMind's framework. And most importantly, it probably won't resemble what science fiction has promised us: neither HAL 9000 nor Skynet. Rather, it will be a continuous increase in capability, the beginnings of which we already see in our daily tools — like Delos Workers, autonomous agents that reliably execute complex workflows, foreshadowing what AGI will make possible on a larger scale.

For businesses, the challenge isn't to wait for AGI idly. It's to choose the right generative AI tools today to boost productivity, while keeping a critical eye on vendors' promises. The security and data sovereignty must remain at the heart of decisions. And for professionals who want to discover the best AI tools of 2025-2026, the best thing to do is still to experiment for themselves.

To delve deeper, we recommend three essential readings:

• Levels of AGI: Operationalizing Progress on the Path to AGI — DeepMind's foundational paper (2023) that defines the 5 levels of AGI.

• Shrinking AGI Timelines: A Review of Expert Forecasts — a rigorous analysis of expert forecasts on the arrival of AGI, by 80,000 Hours (March 2025).

• Artificial General Intelligence in 2026 — a comprehensive overview of AGI's capabilities, economic impact, and challenges in 2026.

AGI is not a question of "if," but of "when" and "how." And it's up to us, collectively — researchers, businesses, citizens — to ensure this path is safe, useful, and shared.

‍