Six months of transformation that can hardly be reduced to a list of models
If you had to pick one sentence to describe the first half of 2026 in artificial intelligence, it would be this: AI has stopped answering and started acting. A paradigm shift that brings enormous opportunities, a fair amount of vertigo, and a political and ethical debate that grows more urgent by the day.
Here is what actually happened.
1. The age of agents: from assistant to autonomous collaborator
The most important development of 2026 is the shift from AI as a response tool to AI as an autonomous executor of complex tasks.
Until recently, the cycle was straightforward: ask a question, get an answer, decide what to do with it. In the first months of 2026, that mechanism was turned upside down. So-called AI agents do not wait for questions : they receive a goal, plan how to reach it, take concrete actions (searching for information, writing code, sending emails, browsing websites, managing files) and report back the results. All without step-by-step supervision.
The spark has a precise name: OpenClaw, launched on January 25, 2026. Austrian developer Peter Steinberger said he built the first version in about an hour. Within weeks it had become one of the fastest-growing open source repositories in GitHub history: an AI agent anyone could run locally, capable of organising files, writing code and browsing the web without routing data through the cloud.
That freedom was both the point and the problem: OpenClaw was accumulating 247,000 GitHub stars while security researchers were calling for it to be shut down. An agent with access to accounts, files and the ability to execute code is enormously powerful in the right hands and catastrophic in the wrong ones.
OpenClaw’s popularity triggered a new race. First came NVIDIA, announcing NemoClaw, an enterprise-grade version with built-in security. Then Anthropic accelerated development of Claude Cowork. OpenAI hired OpenClaw’s founder Peter Steinberger in February 2026 to lead development of their proprietary agentic products.
Within a few months, the question circulating in the industry was no longer “how intelligent is the model?” but “how much can we trust an AI that acts autonomously?” Industry projections suggest that 40% of enterprise applications will incorporate autonomous agents by the end of 2026 (a figure that was practically zero just two years ago).
2. The model that could not be released: the story of Mythos and Fable
The most emblematic episode of the half-year (the one that best illustrates where we have arrived) involves Anthropic and a model called Mythos.
In April, Anthropic unveiled Claude Mythos Preview, a frontier model built for cybersecurity work, simultaneously launching Project Glasswing: a restricted programme giving access to a small group of major technology companies and critical infrastructure providers for defensive security tasks.
The reason for the secrecy was as simple as it was alarming: Mythos had identified flaws in every major operating system and web browser it had been tested on. Rather than releasing it publicly, Anthropic shared it with around 50 vetted organisations (including Amazon, Apple, Google, Microsoft and CrowdStrike) for use exclusively in defensive cybersecurity work. Mozilla alone reported resolving hundreds of vulnerabilities directly through Mythos Preview.
For the first time in AI history, a major lab completed a frontier model and deliberately chose to keep it off the market for safety reasons. A precedent with lasting consequences.
On June 9, 2026, Anthropic released Claude Fable 5, the first Mythos-class model made available to the public, with safeguards blocking responses in high-risk areas such as offensive cybersecurity and biology. The name is deliberate: Fable comes from the Latin fabula, which echoes the Greek mythos (the same story, now tellable in public).
Just days later, the US government ordered Anthropic to withdraw the models. TechCrunch noted the irony: the company had insisted so forcefully on the model’s risks that the US government decided to step in.
A model too powerful to release freely, a “tamed” version for the public, a government that feels entitled to block it (this sequence captures the structural tensions of 2026 better than any other single event).
3. The model wars: speed, parity and the end of easy choices
The competition between models has not let up. The difference from the past, however, is the absence of a clear and stable winner.
GPT-5.4, Claude Opus 4.7 and Gemini 3.1 Pro are all competitive across most benchmarks. Choosing a model in 2026 is primarily a decision about integration into a technology stack, not a decision about raw capability. Put simply, “the best one, full stop” has ceased to exist: there are models better suited to certain tasks, certain ecosystems, certain budgets.
April was the most chaotic month: GPT-5.5 on the 23rd, DeepSeek V4 the day after, Claude Opus 4.7 on the 16th. Gemini 3.1 Pro, Meta’s new model family, Qwen 3, Gemma 4 (all within the same six-week window).
DeepSeek deserves its own paragraph. The Chinese company released V4 with 1.6 trillion parameters, an MIT open source licence and aggressive pricing. Built on Huawei chips (with no dependency on NVIDIA) it demonstrated that China is building an AI infrastructure independent of the West. The Stanford AI Index 2026 confirms it: the gap between the top American and top Chinese model has narrowed to 2.7%.
In May, Google I/O presented Gemini 3.5 Flash as Google’s bet on agentic AI: the model can autonomously execute coding pipelines, manage research projects, and in internal tests built an entire operating system from scratch. Google explicitly declared the end of the chatbot era and the beginning of the agent era.
4. The money: numbers that redefine scale
To understand how seriously the industrial and political world is taking AI, the numbers say it all.
The Trump administration launched Project Stargate, a $500 billion initiative for American AI infrastructure, backed by OpenAI, Oracle, SoftBank and the UAE sovereign wealth fund. The largest investment programme in a single technology in United States history.
Q1 2026 saw more large corporate deals in the AI sector than any previous quarter, with 22 transactions exceeding $10 billion. Meta signed a $100 billion agreement with AMD to diversify its infrastructure and reduce dependence on NVIDIA.
The four major hyperscalers (Google, Amazon, Meta and Microsoft) projected combined capital expenditure of between $600 and $700 billion for 2026, a substantial share earmarked for AI infrastructure.
5. Who governs all this?
As the technology accelerates, the global regulatory landscape has fractured into three mutually incompatible approaches.
In the US, the Trump administration has chosen deregulation and market primacy, betting that slowing innovation would mean ceding ground to China. Yet, as the Mythos case demonstrated, the government still reserves the right to intervene when the stakes are high enough.
In Europe, the AI Act is operational: requirements for high-risk AI systems came into force in January. The practical upshot is that building an AI application that works everywhere in the world now means rewriting its core logic at least three times to comply with different laws.
In China, the government maintains a dual track: maximum support for development, maximum control over public-facing content. With projected investments of $70 billion in 2026 alone and a form of state governance that has no equivalent elsewhere.
The most striking consequence of this fragmentation? Some 12.3% of private AI investment is already going into regulatory compliance tools. AI governance has itself become a billion-dollar market.
6. The impact on work: a change already underway
Anthropic engineers using AI tools produce eight times more code per quarter than their counterparts did in the 2021–2025 period. AI now writes roughly 46% of the average developer’s code, peaking at 61% in languages like Java.
These are current figures, not forecasts. And they extend well beyond coding: LinkedIn’s Labour Market Report 2026 estimates that at least 1.3 million AI-related jobs have been created in the past two years (data annotators, AI engineers, forward-deployed engineers) roles that simply did not exist five years ago.
The paradox of 2026 is this: AI is eroding certain categories of repetitive work while simultaneously opening up new professions that require the ability to direct, evaluate and govern AI systems themselves. The relationship between person and AI is shifting from collaboration to supervision: the worker becomes a “human-in-the-loop governor” (someone who oversees the system rather than executes the tasks).
What to take away from these six months
More than individual models or corporate announcements, the first half of 2026 has delivered some structural certainties.
AI acts, it does not just respond. The shift to autonomous agents has changed the category, not just the feature set. Managing an agent that operates independently on your own systems demands an entirely new kind of accountability.
Power raises concrete ethical dilemmas. The story of Mythos (a model so capable that it required a restricted programme to be deployed safely) will repeat itself. The industry is learning in real time where the limits lie.
The competitive advantage has shifted. Having access to the most powerful model matters less than how you integrate it, govern it, and use it safely. The race for raw capability is becoming a race for infrastructure and governance.
AI geopolitics is a concrete reality. The US, Europe and China hold incompatible visions of how AI should be developed, controlled and distributed. Every company operating globally is already navigating three distinct regulatory regimes today.
The second half of 2026 opens with more powerful models on the way, autonomous agents already in production, and a political and social conversation that is struggling to keep pace with the technology. The most interesting (and most delicate) moment is the one we are living through right now.
