The capabilities of AI double every 7 months: what the METR study reveals

The METR measure, often dubbed the "law of Moore for AI" by analogy with the biennial doubling of transistor density theorized by Gordon Moore in 1965, deserves attention. Not to succumb to the media frenzy accompanying every advancement in the sector, but because it has very concrete consequences on how companies should think about their AI strategy today.

What METR really measures

The uniqueness of the METR study lies in its unit of measurement. Rather than comparing models on classic academic benchmarks, often saturated within months by the latest systems, researchers introduced a metric called time horizon: the duration of a task that a qualified human would take to complete, and that an AI model can now perform with an equivalent success rate.

In plain terms, if a senior developer takes two hours to write a given software module, and an AI model manages to produce this same module with comparable quality, then that model has crossed the “two-hour time horizon.” METR has measured this duration over more than six years, systematically evaluating available models on real software engineering tasks.

The result is clear: since 2019, this duration has doubled approximately every seven months. GPT-2 handled tasks lasting a few seconds. Today, the best models tackle tasks that take several hours for an experienced engineer. The January 2026 update, published by METR under the name Time Horizon 1.1, confirms that this trend is not wavering; it continues, with more data and refined methodology.

You can view the original study directly on the METR website.

Why this measure changes everything

Most leaders perceive the evolution of AI through a news feed: a new model is released by OpenAI, another by Anthropic, Mistral publishes an open source version, Google responds. It’s a continuous stream of novelties that creates the impression of rapid but indistinct progress. The METR curve puts a number on this impression, and that number is staggering.

A doubling every seven months means that in eighteen months, models will be capable of performing tasks about six times longer than today. Over three years, the gap is thirty-fold. In other words, a model struggling today with a complex client file will easily handle the equivalent of a week’s work of an analyst by 2028. This does not mean it will replace the analyst; the question of human added value is much more subtle, but that the range of automatable tasks will expand at a pace that renders any fixed AI strategy obsolete.

This is where thinking shifts from the technical realm to the strategic realm. If model capabilities evolve as quickly, the worst decision a company can make in 2026 is to lock itself into a single vendor. Not because that vendor may be bad today, but because the best model at the moment will, statistically, have been surpassed within six to twelve months.

The real risk is not the model. It is the dependency.

Many organizations, enticed by the ease of integration of ChatGPT Enterprise or Copilot, have built their first AI use cases around a single provider. This was rational two years ago when OpenAI dominated without competition. It is much less so today, in a landscape where Anthropic regularly surpasses GPT in reasoning, Mistral is catching up in multilingual capabilities, Google DeepMind leads in video, and Chinese players like DeepSeek release open-source models on par with the best Western proprietaries.

The pace of progress measured by METR makes this diversity structural. No player can guarantee they will stay ahead for more than a few quarters. A company that has built its entire AI framework on a single engine faces a painful decision every six months: fall behind on capabilities or migrate everything.

The good news is that there is now an architectural response to this issue. Instead of choosing a model, you choose an orchestration layer that gives access to all. This is the philosophy on which we built OUPI: a unique platform that federates over sixty models (GPT, Claude, Gemini, Mistral, Llama, DeepSeek, Flux, Veo, Kling...) with a common memory, shared business integrations, and a unified security framework. When a new model comes out, it simply joins the catalog. Your users do not switch tools, your IT department does not redo integrations, and your compliance remains stable. You follow the METR curve without experiencing abrupt jolts.

Sovereignty is not a hindrance, but an accelerator

One counter-argument often comes up: “If AI is evolving so fast, it’s better to take the American leader and follow the movement, rather than complicating things with sovereignty.” This reasoning overlooks two realities.

The first is regulatory. The European AI Act, now in force, imposes obligations for transparency, traceability, and compliance that cannot merely rely on a commercial contract with a vendor outside the EU. Companies dealing with sensitive data—finance, health, legal, public sector—can no longer afford ambiguity about processing locations, logs, and auditing possibilities.

The second is industrial. Using American models is not incompatible with sovereignty as long as orchestration, memory, storage, and governance remain under European control. This is the stance we have taken at OUPI: the best global models, accessible from a platform hosted in France, with a Privacy Protection mode that guarantees exclusively European routing when the nature of the data requires it. In this approach, sovereignty does not mean forgoing innovation. It means keeping control over the framework in which that innovation is consumed.

What companies should do right now

If we take METR's measure seriously, three convictions follow naturally. First, consider AI as a capability that will continue to expand rapidly, not as a fixed set of tools that we have explored. Second, avoid any architecture that would make the replacement of one model by another costly; this is the very principle of orchestration. Finally, invest in the layers that endure: the organization’s memory, its business integrations, data governance, and usage culture. These layers will outlast all models that come and go.

The curve published by METR is not a prediction. It is an empirical observation over six years of data. It may slow down, accelerate, or break. But betting on its maintenance is today the rational hypothesis, and this is the one on which to build a corporate AI strategy in 2026.