Raedan Institute logo

The State of Artificial Intelligence in May 2026: What Is Happening, What It Means, and What Comes Next

The State of Artificial Intelligence in May 2026: What Is Happening, What It Means, and What Comes Next


Artificial intelligence is moving faster than any technology in human history. This is not a marketing claim. It is a measured assessment that follows from looking at what AI systems could do in 2020, in 2022, in 2024, and comparing that trajectory to what they can do now. The pace of improvement has not been linear. It has been exponential in capability terms, and the social, economic and ethical implications of that trajectory are only beginning to be seriously reckoned with by the institutions — governments, corporations, universities, civil society — that will need to govern this technology's development and deployment.

This piece offers a clear account of where artificial intelligence stands in May 2026: the frontier model developments, the regulatory landscape in the UK and EU, the emerging applications, and the legitimate concerns that sit alongside the undeniable achievements.


Frontier Models: What They Can Now Do

The frontier of AI capability is defined by what the most advanced models can do on the hardest tasks. In 2024 and into 2025, the story was reasoning models — AI systems that could solve complex mathematical problems, write and debug sophisticated code, and navigate multi-step logical challenges at levels that had seemed years away.

OpenAI's o3 model, announced in December 2024, achieved 87.5% on ARC-AGI-1 — the benchmark designed by François Chollet to test flexible, general-purpose reasoning in ways that simple pattern-matching cannot solve (Chollet, 2019). For context, GPT-4o had achieved approximately 5% on the same benchmark. The same model achieved 96.7% on AIME 2024, the American Invitational Mathematics Examination that regularly defeats doctoral students in mathematics (OpenAI, 2024). On SWE-bench Verified — a benchmark that tests AI systems' ability to resolve real software engineering bugs from open-source repositories — o3 achieved 71.7% (Jimenez et al., 2024).

Google DeepMind's Gemini 2.5 Pro has competed at the frontier on different benchmarks, particularly in multimodal reasoning — tasks combining text, images, code and structured data. Anthropic's Claude 4 family, on the trajectory established by Claude 3.5 Sonnet in mid-2024 and Claude 3.7 Sonnet in early 2025, has continued to push capabilities in extended reasoning and reliable instruction-following. The landscape is genuinely competitive, with no single company holding a commanding lead across all relevant benchmarks.

What these benchmark improvements mean in practice is that AI systems in 2026 can write professional-quality code, draft complex legal and financial documents, conduct research literature reviews, engage in sophisticated mathematical reasoning, and perform many creative and analytical tasks at levels that were, until very recently, the exclusive domain of highly educated human professionals. This is not a hypothetical future capability. It is the current state of affairs.


The DeepSeek Moment and Its Implications

The most significant development of early 2025 was not from the established Western AI laboratories. It was from DeepSeek, a Chinese AI company, whose R1 model — released under a permissive MIT open-source licence in January 2025 — reportedly achieved performance comparable to OpenAI's o1 at a fraction of the training cost, with estimates suggesting the training run cost approximately $5.6 million compared to hundreds of millions for comparable Western models (DeepSeek-AI, 2025).

The market reaction was immediate and dramatic. NVIDIA lost over $600 billion in market capitalisation in a single day as investors recalibrated their assumptions about the compute intensity required for frontier AI development. The episode demonstrated that the assumption underpinning Western export controls on advanced semiconductors — that restricting access to chips would restrict access to frontier AI — was more fragile than policymakers had assumed. If frontier AI can be achieved with significantly less compute than previously believed, through algorithmic efficiency rather than raw scale, the strategic calculus around hardware export controls becomes considerably more complicated.


Agentic AI: From Assistants to Agents

The most consequential shift in AI capability in 2025 and 2026 has been the emergence of genuinely capable agentic systems — AI that does not merely respond to prompts but takes sequences of actions over time to accomplish complex goals. OpenAI's Operator, Anthropic's computer use capability and Google's Project Mariner all demonstrated AI systems that could operate computers, browse the internet, complete multi-step tasks across applications, and accomplish goals specified in natural language without step-by-step human instruction.

The implications of this shift are profound. AI that can merely answer questions is a very useful tool. AI that can independently take actions — book appointments, file documents, conduct research, send communications, manage workflows — begins to function as something more like a colleague or an employee. This represents a qualitative change in the economic and social implications of AI, and it is happening much faster than most governance frameworks anticipated.

The safety implications of agentic AI are significant. An AI system that takes actions in the world can cause harm — through errors, through misunderstanding of intent, through being deceived by malicious content in the environment it is operating in. The research community has been working on these problems under the heading of "agent safety", but the deployment of agentic systems has in many cases outpaced the maturity of the safety frameworks available to govern them.


The Regulatory Landscape: UK, EU and US in 2026

The European Union's Artificial Intelligence Act — the world's first comprehensive AI regulation — entered its most significant enforcement phase in August 2026, when obligations on general-purpose AI model providers including OpenAI, Anthropic, Google and Mistral came into full effect. The Act classifies AI systems by risk level — unacceptable, high, limited and minimal — and applies graduated requirements accordingly. Violations carry fines of up to €35 million or 7% of global annual turnover for the most serious infringements (European Parliament, 2024).

The UK has taken a different approach. The AI Safety Institute — renamed the AI Security Institute in February 2025 to reflect a broader national security focus — has evaluated frontier models before and after release, publishing technical reports that are available to policymakers and the public. The AI Opportunities Action Plan, published by HM Government in January 2025 and authored by Matt Clifford CBE, set out fifty recommendations for accelerating AI adoption in the UK economy, including the establishment of AI Growth Zones to accelerate computing infrastructure, a twenty-fold expansion of public AI compute by 2030, and a National Data Library to make public sector data available for AI development (HM Government, 2025).

The UK has deliberately avoided the EU's prescriptive regulatory approach, preferring to work through existing regulators — the ICO, the FCA, the CMA — applying sector-specific guidance to AI applications within their remit. Whether this lighter-touch approach will prove adequate as AI capabilities advance, and whether the UK can attract the AI investment it seeks while the EU's regulatory framework creates compliance costs that disadvantage European-based development, are questions that will be answered in the next few years of this technology's development.


The Stargate Investment and the Scale of the Race

In January 2025, President Trump announced the Stargate Joint Venture — a commitment of $500 billion in AI infrastructure investment over four years, involving OpenAI, Oracle, SoftBank and the Abu Dhabi sovereign wealth vehicle MGX (The White House, 2025). The scale of this commitment — equivalent to the entire annual GDP of a medium-sized European economy — reflects the degree to which AI has become a matter of national strategic priority for the United States, framed explicitly in terms of competition with China.

The race dynamic in AI development — between companies, between nations, between regulatory philosophies — is one of the most significant structural features of the current moment. It creates pressure to deploy systems quickly, to prioritise capability development over safety research, and to treat regulatory compliance as a constraint to be minimised rather than a legitimate social good. These pressures are real, and the institutions working on AI safety and governance are working against them. The outcome of that contest — whether AI develops in ways that are safe, beneficial and equitably distributed — is among the most consequential open questions of the twenty-first century.


References

  • Chollet, F. (2019) 'On the Measure of Intelligence', arXiv preprint, arXiv:1911.01547.
  • DeepSeek-AI (2025) 'DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning', arXiv preprint, arXiv:2501.12948.
  • European Parliament (2024) Regulation (EU) 2024/1689 of the European Parliament and of the Council: The AI Act. Brussels: European Parliament.
  • HM Government (2025) AI Opportunities Action Plan. London: Department for Science, Innovation and Technology.
  • Jimenez, C.E. et al. (2024) 'SWE-bench: Can Language Models Resolve Real-World GitHub Issues?', arXiv preprint, arXiv:2310.06770.
  • OpenAI (2024) OpenAI o3 System Card. San Francisco: OpenAI.
  • The White House (2025) Stargate Joint Venture Announcement, 21 January. Washington DC: The White House.