For most people, AI still means a chat window: you type a question, it types an answer. But the most capable AI models are increasingly being built and sold as something else. They plan, use tools, check their own work and carry out a job across many steps, sometimes for hours. This guide explains what has changed, the terms you will keep hearing, and how to judge the claims.
From answering to doing
The first wave of large language models were, at heart, very good text predictors. You gave them a prompt and they produced the most plausible continuation. That made them useful for drafting, summarising and explaining, but they could not look anything up, run a calculation reliably or take an action in the world.
Today's frontier models, meaning the most capable systems from the leading labs, are wrapped in software that lets them do far more. Three ingredients matter most.
1. Reasoning
Reasoning models spend extra effort thinking through a problem before they answer, breaking it into steps and checking intermediate results. Many products now let you choose an effort level: low for quick replies, high or maximum for hard problems. More effort usually means better answers, but also more cost and a longer wait.
2. Tool use
Tool use (sometimes called tool calling or function calling) lets a model hand part of a job to other software. Instead of guessing a figure, it can run code, search the web, query a database or open a spreadsheet. The model decides which tool to use, reads the result and carries on. This is what turns a chatbot into something that can produce a working file or a checked calculation.
3. Agents
An agent is a model given a goal, a set of tools and permission to work through many steps on its own. Google's Gemini Tasks, for example, can now book appointments, analyse bills and buy items with the user's approval, working inside an isolated virtual machine. Coding agents can edit a whole codebase, run its tests and fix what breaks. We look at agents in detail in AI Agents in Practice.
Why size is no longer the headline
A few years ago, launches were measured by parameter counts: the number of adjustable values inside a model. Bigger generally meant smarter. That number is now less useful for buyers, for three reasons.
First, labs often do not publish it. Second, much of the improvement now comes from training methods, reasoning time and the tools around the model rather than raw size. Third, what organisations actually pay for is finished work, and that depends on how many tokens (small chunks of text) a model uses to complete a task, not how large it is.
That is why launch announcements increasingly lead with efficiency. When OpenAI released GPT-6.2 Astra and Sol in November 2026, it emphasised that both models finish demanding tasks with fewer output tokens, and the independent group Artificial Analysis now reports a cost per task alongside its intelligence scores. A model that is slightly less clever but uses far fewer tokens can be the better buy.
Reliability is the new battleground
When a chatbot gets something wrong, you usually see it straight away. When a multi-step system gets something wrong at step four of twenty, the error can be buried in a finished-looking report. So the question that matters most is not "how smart is it?" but "how often does it get the whole job right, and does it tell me when it can't?"
Two examples from recent launches show why. Artificial Analysis measures how often models guess rather than admit they do not know an answer; it reported hallucination rates of 59% and 55% for OpenAI's two new models at maximum effort. And OpenAI's own testing found those models were more likely than their predecessors to leave out obstacles that blocked their work, such as a missing file, something the company says it is working to reduce.
Neither finding means the models are bad. They mean that benchmark scores and real-world dependability are different things, and that systems which act on your behalf need checking in proportion to what they can do.
What the benchmarks do and don't tell you
Benchmarks are standardised tests. Some measure knowledge, some coding, some long professional tasks. They are useful for spotting broad trends, but they have limits:
- Company-reported scores are chosen by the company. Look for independent results as well.
- Different tests move in different directions. A new model can rise on one index and fall on another.
- Your tasks are not the benchmark's tasks. A model that tops a legal test may not suit your firm's documents.
Our buyer and builder checklist explains how to run your own tests.
The UK angle
The UK does not build its own frontier model and, according to the AI minister, does not plan to fund one, relying instead on trusted access to models made by allies. That makes testing and oversight central to Britain's role.
The AI Security Institute (AISI), part of the Department for Science, Innovation and Technology, describes itself as the first state-backed organisation dedicated to making advanced AI safe, secure and beneficial. It researches what advanced models can do, evaluates risks to national security and public safety, and works with developers and other governments. Its recent research has compared how far open-weight models, whose internals anyone can download, trail closed frontier models on cyber tasks.
The shift to multi-step systems also changes the security picture. The National Cyber Security Centre (NCSC) warned in its November 2026 annual review that AI adoption by hostile actors is highly likely to increase the scale and efficiency of attacks within two years, and that AI-powered malware is highly likely to be used in a ransomware attack on UK critical national infrastructure by 2028. The same capabilities that let an AI agent fix code can help an attacker find weaknesses faster.
For UK organisations, the practical upshot is simple: assume your AI will come from overseas labs, choose it on cost per task and reliability rather than hype, and put the same care into controlling what it can access as you would for a new member of staff.
Key terms at a glance
- Frontier model: one of the most capable AI models available at a given time.
- Token: a small chunk of text that models read and write; usage is usually priced per million tokens.
- Reasoning effort: how long a model thinks before answering; higher effort costs more.
- Agent: a model given a goal and tools and allowed to work through many steps itself.
- Hallucination: a confident answer that is wrong or made up.
FAQs
What is the difference between a chatbot and an AI agent?
A chatbot answers a message. An agent is given a goal, tools and permissions, and works through many steps on its own, such as searching, running code or making a booking, before reporting back.
Why do AI companies now talk about cost per task?
Because organisations pay for finished work. A model that uses fewer tokens to complete a job can be cheaper overall, even if its price per token is similar or its benchmark scores are slightly lower.
Does the UK have its own frontier AI model?
No. The UK relies on models built by allies, and its AI Security Institute focuses on testing and understanding the risks of advanced AI rather than building it.
Sources
- OpenAI: Introducing GPT-6.2 Astra and Sol
- Artificial Analysis: GPT-6.2 Astra and Sol results
- Google: Put Gemini to work for you
- UK AI Security Institute
- NCSC: UK facing AI-powered attacks on critical infrastructure within two years
- City AM: UK does not need its own frontier AI model, says AI minister
Follow AI Decoded on Substack
Weekly plain-English AI news and verified UK events.
Subscribe