Skip to main content

How AI Assistants Work in 2026

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

You do not need the history of artificial intelligence to use an assistant well. You need an accurate mental model of the thing in front of you: what it is made of, what it can reach, what it remembers, and where it tends to be wrong. This lesson gives you that model in one pass, with the 2024-era assumptions removed.

Learning focus​

  • Understand the difference between the model and the product wrapped around it
  • Know what tokens and a million-token context window mean for your work
  • Use the effort setting instead of asking the assistant to "think step by step"
  • Know what the assistant can see (web, files, memory, your accounts) and where its errors cluster

A model wrapped in a product​

Underneath every assistant is a large language model (LLM): a system trained on an enormous amount of text and other data that learned to produce the most plausible continuation of whatever it is given. That single skill, generalized across billions of examples, is what lets it draft, summarize, translate, reason through a problem, and write code. The model has no live connection to anything. On its own it only knows what was in its training data.

The assistant you use is the model plus a product around it. The product adds web search, file reading, code execution, memory across conversations, voice, and connectors into your mail, calendar, and documents. In 2026 the product also adds agent modes, in which the assistant plans a multi-step task and carries it out through those connectors, pausing for your approval at sensitive steps. When people say "the AI can't browse" or "it forgets everything between chats," they are describing the bare model, not the product. Those statements were true in 2023 and are false for the mainstream apps today.

Machine learning and deep learning are the techniques used to train these models; you will meet the terms, but you do not need them to use the tools. The AI Glossary has short definitions when you want them.

Tokens and the million-token context​

Models read and write in tokens, pieces of text roughly three-quarters of a word long in English. Everything in a conversation counts toward the context window, the total amount the model can hold at once: your messages, its replies, any files you attach, and the hidden instructions the product adds. In 2026 the leading models from Anthropic, OpenAI, and Google all offer context windows of about one million tokens. On Anthropic's current tokenizer, one million tokens is roughly 555,000 words, enough for several books or a year of meeting notes.

That capacity changed the advice. Splitting a long report into chunks, or summarizing the conversation every hour so the model does not "forget," is no longer necessary for ordinary work. Attach the whole document and ask your question.

Two limits survive. First, long context is not free of errors. A March 2026 study of 35 open-weight models found that the best of them fabricated about 1% of answers when the document fit in 32,000 tokens but could not stay under 10% at 200,000 tokens (arXiv 2603.08274). The lesson is to attach the relevant files, not every file you own. Second, a long, meandering conversation accumulates stale instructions and half-corrected drafts. When a thread has gone wrong twice, a fresh conversation with a better brief usually beats a third correction; the Check lesson explains when to restart.

Reasoning is always on; effort is the dial​

In 2023 you could improve answers by typing "think step by step." That trick is obsolete. The current flagship models reason internally before every answer, and on Anthropic's top models thinking cannot be turned off at all. What you control instead is effort: how much reasoning, searching, and tool use the model spends on a request. Anthropic exposes effort levels from low to max; OpenAI and Google expose equivalent reasoning or thinking levels. The default is a middle setting that suits most tasks.

Two practical consequences follow. At low effort a model is more likely to answer from memory than to run a search or open a file, so if currency or sources matter, raise the effort or ask it explicitly to search. And adding ritual phrases now does little or nothing. Wharton's Prompting Science research found that chain-of-thought prompts gave reasoning models only marginal accuracy gains while costing 20-80% more time (arXiv 2506.07142), and a March 2026 medical preprint found the phrase could lower accuracy (arXiv 2603.25960). Spend your words on information the model does not have, not on instructions about how to think. The Decide lesson covers how to choose a mode and effort level.

Search is on; the cutoff is months old, not years​

Every model has a knowledge cutoff, the date after which its training data ends. In 2026 that lag is months: the Claude models released in September and October 2026 have a stated cutoff of June 2026. More importantly, the mainstream apps search the web by default when a question calls for it, including Claude's free tier. "The AI doesn't know recent events" is therefore the wrong worry. The right one is whether, for this request, the model searched or answered from memory. If you need a current figure, ask for a search and for the source links, then open them.

Memory across conversations​

The assistant can now remember you between sessions. ChatGPT, Claude, and Gemini each offer a memory feature that stores facts and preferences from past chats and brings them into new ones; Claude includes it on the free tier, and Anthropic reported in August 2026 that memory from chat and from agent sessions was merged into one record. Projects, where you group related conversations and files, typically keep their own memory. Each app also offers a temporary or incognito chat that is not remembered, and in the EU and UK memory is off by default in ChatGPT.

Memory is powerful and easy to pollute. The Supply Context lesson shows how to review and edit what the assistant has stored, and the Protect lesson covers what should never go in.

What it can take in and put out​

The current models accept text and images as input, and the apps add voice, file upload, and document creation. ChatGPT and Gemini generate images and, in Gemini's case, video inside the chat; Claude's models produce text and files but do not generate images, audio, or video natively. All three can write and run code, which is how they build spreadsheets, charts, and small apps on request. Through connectors and agent modes they can also act: read your inbox, add a calendar event, edit a document in your drive. The Make Things pages cover each output type; the Agents in the Apps You Already Pay For page covers acting.

Where it fails​

Assistants still produce confident, fluent errors. What changed is that the error rate is now clearly task-dependent, so "verify everything" can become "know where to look."

Grounded tasks, such as summarizing a document you attached, are the most reliable; the best models fabricate in the low single digits there. Errors climb for specific citations, exact figures, long-context recall, and legal or medical specifics. The clearest evidence is in courtrooms: trackers counted roughly 2,000 US cases involving fabricated AI citations by September 2026, up from about 200 a year earlier (reported, secondary), and judges have sanctioned and suspended lawyers over them. In code, a May 2026 study found 4.6-6.1% of package names suggested by five frontier models did not exist (arXiv 2605.17062).

Two newer failure modes come from the models being better at following instructions. They are literal: ask "can you suggest some changes" and you may get suggestions rather than the change you wanted. And they overreact to aggressive emphasis; Anthropic's own guidance now tells users to dial back "CRITICAL" and "MUST" language written for older models. Finally, models reflect the biases in their training data, so default assumptions about who a nurse or a CEO is still show up and still need catching. The Check lesson turns all of this into a short routine.

Key takeaways​

  • The assistant is a model plus a product; the product searches, reads files, remembers, and acts through connectors
  • Million-token context means you attach whole documents, but relevant files beat all files, and long threads degrade
  • Reasoning is built in; choose an effort level instead of typing "think step by step"
  • Web search is on by default and cutoffs are months old, so the question is whether it searched this time
  • Errors cluster around citations, exact figures, long-context recall, and legal or medical specifics

Next steps​

Continue to: Decide: What to Delegate, Which Tool, Which Mode

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks