What Has Actually Changed in Everyday AI Tools
A caution before anything else. Anything written about specific AI model versions dates within months, sometimes weeks. Version numbers, pricing tiers and feature names change constantly, and any article that hangs its advice on a particular release will be wrong shortly after publication. This piece deliberately talks about capability categories instead, because those move slowly enough to plan around.
The Four Capability Shifts That Actually Matter
Strip away the marketing and four things have genuinely changed in the assistants most professionals use, namely Claude, ChatGPT, Gemini and Copilot.
- Long context. Context windows have grown from a few thousand words to the equivalent of entire books. Several major assistants now advertise windows measured in the hundreds of thousands of tokens, and in some cases around a million.
- File handling. Uploading a PDF, spreadsheet, image or set of documents and asking questions across them is now standard rather than a premium extra.
- Live web access. All of the mainstream assistants can retrieve current web content rather than relying solely on training data, which changes what kinds of questions are safe to ask.
- Tool use. Assistants can now call external tools and systems rather than only producing text. This is the largest of the four shifts, and the least understood.
Long Context Is Real, But Advertised Is Not the Same as Reliable
The practical effect of long context is significant. You can put a full policy document, a long contract, a set of meeting minutes or an entire report into a single session and ask questions across the whole thing, rather than chopping it into fragments and losing the thread between them.
The caveat is that an advertised context window describes how much text a model will accept, not how reliably it will use every part of it. Recall across a very long input is uneven, and quality tends to degrade well before the stated limit. Independent evaluations of long-context recall exist, but the numbers differ by test design and change with every release, so treat any single figure with suspicion.
The working habit that follows is simple: give the model the whole document if you have it, but ask targeted questions rather than open-ended ones, and check anything load-bearing against the source yourself. Long context reduces the amount of copying and pasting. It does not remove the need to verify.
From Answering to Doing
The most consequential change is the shift from assistants that produce text to assistants that take actions. This is usually labelled agentic, which is a vague word for a concrete thing: the model can call an external tool, receive the result, and decide what to do next.
The infrastructure behind this converged faster than most people noticed. The Model Context Protocol, an open standard for connecting AI assistants to external tools and data sources, was released by Anthropic in late 2024. Within a year it had been adopted well beyond its origin, with support across OpenAI, Google, Microsoft and AWS products, and in December 2025 it became a founding project of the Agentic AI Foundation, a fund under the Linux Foundation. Standards wars in this industry usually take longer and end less cleanly.
For a professional user, the practical meaning is that an assistant can increasingly reach your actual working systems, calendars, files, ticketing tools, code repositories, rather than sitting in a chat window while you shuttle information back and forth. Assistants that operate a browser or a desktop directly have also moved from demonstration to limited release.
This is where the risk profile changes. A text assistant that gets something wrong produces a bad draft you can discard. A tool-using assistant that gets something wrong takes an action in a real system. The correct posture is to grant narrow permissions, keep a human approval step for anything consequential, and be sceptical of any workflow that removes the review point in the name of efficiency.
Where the Four Big Assistants Sit
Rather than rank them, which would be out of date immediately, it is more useful to note what tends to drive the choice.
- Ecosystem is the dominant factor for most organisations. Copilot's advantage is that it is embedded in Microsoft 365, and Gemini's is that it is embedded in Google Workspace. If your documents and mail already live in one of those, the integration usually outweighs marginal model differences.
- Long-document and careful-writing work is where Claude has built its reputation, and where its context handling is most often cited.
- Breadth of third-party integrations and image generation remains a ChatGPT strength.
- Data governance is frequently the deciding factor in regulated sectors, and it has nothing to do with model quality. Where data is processed, whether it is used for training, and what the enterprise terms say will determine what you are permitted to use regardless of which tool you prefer.
The Rules Have Moved Too
If you work in or with the EU, the regulatory position has shifted alongside the technology. Obligations on providers of general-purpose AI models under the EU AI Act applied from 2 August 2025. The high-risk obligations originally due in August 2026 were subsequently deferred under the Digital Omnibus package agreed in 2026, with standalone high-risk systems in the employment and similar categories pushed to December 2027.
The point is not the dates themselves, which may move again. It is that if you are deploying AI in recruitment, education, credit or any other sensitive area, the compliance position is a live question with a moving timetable, and it should be checked rather than assumed.
What To Do About Any Of This
The skill that transfers across every release is not knowing which model is currently best. It is knowing how to specify a task precisely, how to supply the right context, how to check the output, and where to draw the line on what you let a tool do unsupervised. Those habits survive every version bump.
Our AI & Digital Skills series covers this ground practically across fifteen titles, including dedicated guides to ChatGPT and Claude, and AI for Teachers for those applying these tools in a school setting.