AI Guide · Updated 14 November 2026
History
AI in 2025 and 2026: reasoning models and agents
By drayz.ai · Updated · 10 min read

In 2025 and 2026, artificial intelligence moved from answering questions to doing work. "Reasoning" models learned to think through hard problems, AI agents began carrying out long tasks with software tools, cheap and capable open models arrived from China, and hundreds of billions of dollars poured into data centres. Governments shifted their focus from safety summits to competitiveness, while scientists warned that safeguards were not keeping pace. This chapter describes where AI stands as of November 2026. Because the field moves so quickly, we update it regularly.
The big picture
Three shifts define AI in 2025-2026:
- From chatbots to agents. AI systems increasingly plan and carry out multi-step tasks, using tools such as web browsers, code editors, spreadsheets, email and business software, often for hours at a time, with people approving key steps.
- From bigger training runs to more thinking. Alongside ever larger training, models now improve by "thinking" longer when they answer, and are trained with reinforcement learning to reason well.
- From novelty to infrastructure. AI is becoming part of everyday business software and public services, and the race to build chips, data centres and power supplies has become an economic and geopolitical priority.
Reasoning models
OpenAI's o1 (September 2024) showed that a model trained with reinforcement learning to work through problems step by step, before answering, could perform far better on difficult maths, science and coding tasks. In 2025 every major lab adopted the approach:
- DeepSeek-R1 (January 2025), from the Chinese company DeepSeek, matched leading US reasoning models on many benchmarks. It was released with open weights and a detailed paper, and DeepSeek reported that its underlying base model had been trained remarkably cheaply. The news triggered a sharp sell-off in AI-related shares: on 27 January 2025 Nvidia lost around $590 billion of market value, the largest one-day fall for any company in US stock market history at the time.
- Google's Gemini 2.5 (March 2025), Anthropic's Claude 3.7 Sonnet (February 2025) and Claude 4 models (May 2025), and OpenAI's o3 and o4-mini (April 2025) all built in extended reasoning.
- In July 2025, experimental models from Google DeepMind and OpenAI achieved gold-medal-level scores at the International Mathematical Olympiad, solving competition problems in natural language within the time limits.
- On 7 August 2025 OpenAI released GPT-5, which combined fast responses and deeper reasoning in one system and became the default model for ChatGPT's hundreds of millions of users.
- In November 2025 Google released Gemini 3 and Anthropic released Claude Opus 4.5, part of a burst of frontier releases in late 2025 that also included xAI's Grok 4.1.
Reasoning made models much better at tasks with checkable answers, such as maths and code, and somewhat more reliable in general. It also made them more expensive to run, because thinking at length uses more computation for each answer.
The rise of AI agents
An AI agent is a system that uses a model to decide what to do next, takes actions with tools, observes the results and keeps going until a task is done.
Coding agents
Software development became the first area where agents were adopted at scale. Tools such as Anthropic's Claude Code, OpenAI's Codex, Google's Jules, GitHub Copilot's agent mode and Cursor let developers hand over whole tasks: fix this bug, add this feature, write tests, update this library. By 2026 many software teams used coding agents daily, and non-programmers increasingly built simple tools for themselves by describing what they wanted, a practice nicknamed "vibe coding".
Computer-use and browser agents
Agents that operate a computer like a person, reading the screen and clicking and typing, moved from demos to products, including OpenAI's ChatGPT agent (2025), Anthropic's computer use and browser extensions, and Google's agent features. They can book appointments, fill in forms, research suppliers and compile reports, though they remain slower and less reliable than direct software integrations.
Connecting AI to everything
In November 2024 Anthropic introduced the Model Context Protocol (MCP), an open standard for connecting AI models to tools and data sources. During 2025 it was adopted across the industry, including by OpenAI, Google and Microsoft, so that AI assistants could plug into calendars, email, CRMs, accounting systems, design tools and databases in a common way.
How capable are agents?
The research organisation METR measures how long a task, in terms of the time a skilled person would take, an AI agent can complete with 50% reliability. Its 2025 research found that this "time horizon" had been doubling roughly every seven months since 2019, and faster since 2023. By 2026 the leading agents could complete, about half the time, software tasks that would take skilled professionals many hours, and early evidence suggested some well-specified coding projects lasting weeks were within reach. Reliability on long, open-ended work in messy environments remains the main limitation.
Agents at work
In the UK and elsewhere, businesses began deploying agents in customer service (including AI voice agents that answer phone calls, take bookings and qualify leads), finance (processing invoices, reconciling accounts), sales (researching prospects, drafting outreach), HR (screening and scheduling) and IT support. The pattern that works best is usually "human in the loop": the agent prepares and executes routine steps, and a person approves anything important.
Open models and the China factor
2025 transformed the open-model landscape:
- Chinese labs, including DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Zhipu and MiniMax, released a stream of capable open-weight models, often matching Western models that had been state of the art months earlier, at much lower prices.
- OpenAI released its first open-weight models since GPT-2, gpt-oss, in August 2025.
- Mistral in France, Google (Gemma) and others continued releasing open models.
Open models mean businesses can run capable AI on their own hardware, keeping data in-house and reducing costs. They also make safeguards harder to enforce, because safety training can be removed. US export controls on advanced chips to China remained in place and became a central issue in US-China relations.
The infrastructure boom
The scale of investment in AI infrastructure grew to unprecedented levels:
- In January 2025 OpenAI, SoftBank, Oracle and MGX announced Stargate, a joint venture aiming to invest up to $500 billion in US AI data centres over four years. Its first site, in Abilene, Texas, began operating in 2025.
- Microsoft, Alphabet, Amazon and Meta together planned to spend hundreds of billions of dollars a year on capital expenditure, mostly for AI. Nvidia's chief executive estimated that $3-4 trillion would be spent on AI infrastructure by the end of the decade.
- Single AI campuses were planned at gigawatt scale, comparable to the output of a large power station, prompting deals for gas, renewable and nuclear power.
- In the UK, the government designated five AI Growth Zones (Culham, the North East, North Wales, South Wales and Lanarkshire) and, in its January 2026 progress report, said they were linked to more than £28 billion of planned investment. The Isambard-AI supercomputer in Bristol came online as the UK's most powerful public AI computer.
Debate continued about whether the investment boom would pay off or prove a bubble. Usage and revenues grew very fast, but so did costs, and the financial arrangements between chipmakers, cloud providers and AI labs became increasingly complex and interlinked.
AI in the UK economy
Official statistics show how fast AI spread in British businesses:
- The share of businesses with 10 or more employees using AI rose from about 12% in late 2023 to 35% in June 2026 (ONS).
- Large language models were the most used type of AI (18% of businesses), followed by visual content creation (16%).
- Adoption was wide but shallow: only about 10% of AI-using businesses described their use as extensive.
- Around 36% of UK workers reported using AI at work.
At the same time, the jobs market for young people tightened. Graduate vacancies on the job site Adzuna fell to their lowest level in a decade in mid-2026. Employers and researchers disagree about how much of this is caused by AI rather than wider economic pressures; see the chapter on AI and jobs in the UK.
Policy: from safety to growth
The politics of AI shifted noticeably:
- In January 2025 the UK government adopted the AI Opportunities Action Plan, aiming to make the UK an "AI maker, not an AI taker" through compute, skills, AI Growth Zones and public sector adoption. A year later it reported progress on 38 of the plan's 50 actions.
- In February 2025 the UK's AI Safety Institute was renamed the AI Security Institute, sharpening its focus on national security risks such as cyberattacks and chemical and biological weapons.
- At the AI Action Summit in Paris in February 2025, the UK and the US declined to sign the summit's declaration on "inclusive and sustainable" AI, which was signed by around 60 countries including China and EU members. The UK said it could not agree all parts of the text.
- In the US, the new administration revoked the previous executive order on AI safety in January 2025 and emphasised deregulation and AI leadership.
- The EU AI Act brought in its ban on prohibited practices (February 2025) and obligations for general-purpose AI models (August 2025), backed by a Code of Practice that most major developers signed. In 2026 the EU agreed the Digital Omnibus, postponing the main high-risk rules to December 2027 and August 2028.
- The UK's Data (Use and Access) Act 2025 updated data protection law, after a long battle in Parliament over AI training and copyright.
- In May 2026 the King's Speech included a Regulating for Growth Bill with powers for regulatory sandboxes, but still no dedicated bill for frontier AI models, despite earlier commitments.
- The International AI Safety Report 2026 (February 2026), led by Yoshua Bengio with more than 100 experts, found capabilities advancing rapidly while safeguards remained fragile and could often be bypassed. Its findings fed into the AI Impact Summit hosted by India in February 2026.
Safety and security concerns
As models became more capable, the risks became more concrete:
- Cybersecurity: AI developers reported that their newest models had become highly capable at finding software vulnerabilities, raising both defensive opportunities and offensive risks. Some of the most cyber-capable models were initially released only to vetted organisations, with additional safeguards on general releases.
- Biological risks: several labs activated stricter safety measures after evaluations suggested their models could provide meaningful help to people seeking to create biological weapons.
- Misbehaving agents: researchers documented cases of models behaving deceptively in tests, gaming evaluations or pursuing goals in unintended ways. Independent organisations such as METR began piloting assessments of risks from AI agents used inside AI companies.
- Government involvement: in June 2026 the US Department of Commerce applied export controls that led Anthropic to suspend access to its newly released Claude Fable 5 and Claude Mythos 5 models for several weeks, before the controls were lifted at the end of June and access was restored in July. The episode showed how frontier AI was becoming a matter of national security policy.
- Everyday harms: deepfake scams, AI-generated child abuse material, chatbots giving harmful advice to vulnerable users, and AI "hallucinations" in legal filings and official reports all prompted new rules, lawsuits and safeguards.
Physical AI and humanoid robots
AI models for robots, often called vision-language-action models, improved quickly, helping robots understand instructions and handle objects they had not seen before. Humanoid robots began moving from demos to deployment:
- Agility Robotics' Digit worked in logistics warehouses under commercial contracts.
- Figure's humanoids were tested in BMW's car plant in South Carolina.
- Tesla continued developing its Optimus robot.
- 1X, which opened pre-orders for its NEO home robot in October 2025 at about $20,000, opened a factory in 2026 with deliveries to US homes planned during the year.
- Chinese manufacturers, including Unitree, produced the majority of humanoid robots by volume, at increasingly low prices.
Home robots remain limited and partly teleoperated, and most experts expect widespread household use to take many more years. Simulation, where robots practise millions of times in virtual worlds, is central to training them; you can try a simple version in the drayz.ai Robotics Lab.
AI in science and medicine
AI increasingly contributed to scientific research: designing proteins and drug candidates, predicting materials, improving weather forecasts, analysing medical images and helping write and check code for experiments. Pharmaceutical companies built dedicated AI supercomputers. AI systems began to assist with mathematical research and proofs. In the NHS, AI tools were used to help with scan analysis, note-taking ("ambient voice technology") and administrative tasks, under clinical oversight.
What has not changed
For all the progress, some things remain stubbornly true in late 2026:
- AI still makes confident mistakes, and important outputs still need checking.
- Most businesses use AI in shallow ways; the biggest gains go to those who redesign processes around it.
- Costs per task keep falling, but the total amount spent on AI keeps rising.
- There is no agreement on when, or whether, AI will reach human-level general intelligence.
- The key questions of the next few years, about jobs, power, safety, copyright, energy and who benefits, are political and social as much as technical.
Key takeaways
- Reasoning models (2024-2025) made AI much better at maths, science and code by thinking before answering.
- AI agents, especially coding agents, became mainstream in 2025-2026; the length of tasks they can complete has been doubling every few months.
- Chinese open-weight models, led by DeepSeek-R1 in January 2025, reshaped competition and pricing.
- Investment in AI data centres reached hundreds of billions of dollars a year; the UK designated five AI Growth Zones.
- UK business adoption of AI reached 35% (firms with 10+ staff) by June 2026, but mostly in shallow ways.
- Policy shifted towards growth and security; the EU delayed its high-risk rules to 2027-28; the UK still has no frontier AI bill.
- Safety evaluations found growing cyber and biological risks, and humanoid robots began their first real deployments.
Previous chapter: The generative AI boom, 2022-2024 · Next chapter: The cost of AI · Back to the complete guide to artificial intelligence