Blog · November 11, 2026

From chat to AI that acts

AI now calls tools, uses a computer, writes code, and carries context forward. This note follows the posts the companies published.

From a reply to a tool call

On June 13, 2023, OpenAI added function calling to the Chat Completions API. Developers can describe functions to gpt-4-0613 and gpt-3.5-turbo-0613. The model can return a JSON object with the arguments for one of those functions.

OpenAI described this as a way to connect GPT with external tools and APIs. A person can ask for the weather, or ask who the top customers are this month. The model names the function. The developer's code runs it. The same post says a developer can also ask the model to pull structured details out of a page of text.

One way to connect many tools

On November 25, 2024, Anthropic open sourced the Model Context Protocol, called MCP. It is an open standard for connecting an assistant to the systems where data lives. A developer can publish data through an MCP server, or build an app that connects to those servers.

That launch included the specification, software kits, local MCP support in the Claude desktop apps, and a public set of servers. The first servers covered Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer.

On December 9, 2025, Anthropic gave MCP to the Agentic AI Foundation, a fund under the Linux Foundation. Anthropic, Block, and OpenAI founded it together. Google, Microsoft, Amazon Web Services, Cloudflare, and Bloomberg support it. MCP sits there with goose from Block and AGENTS.md from OpenAI.

In that same post, Anthropic said there were more than 10,000 active public MCP servers. It said ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code had adopted MCP. Claude's directory had more than 75 connectors. The Python and TypeScript kits had more than 97 million downloads a month.

AI that uses a screen

On October 22, 2024, Anthropic released computer use in public beta. A developer can ask Claude to look at a screen, move a cursor, click, and type. It shipped on the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI.

In December 2024, Google introduced Project Mariner with Gemini 2.0. It is a research prototype that reads a browser screen, including text, code, images, and forms, and acts through a Chrome extension. Google said it can type, scroll, and click in the active tab, and that it asks before sensitive steps such as a purchase. On the WebVoyager test of real web tasks, Google reported 83.5% for Mariner as a single AI.

On May 20, 2025, Google said Mariner could run up to ten tasks at once, such as looking things up, booking, buying, and research. That version was available to Google AI Ultra subscribers in the United States.

On January 23, 2025, OpenAI released Operator, a research preview for Pro users in the United States. Operator uses its own browser. It looks at a page and types, clicks, and scrolls. The model is called Computer-Using Agent. On July 17, 2025, OpenAI folded Operator into ChatGPT as ChatGPT agent. In that mode, ChatGPT uses its own virtual computer. The tools named in the post are a visual browser, a text browser, a terminal, and direct API access. ChatGPT asks before actions that matter, and a person can stop the task or take the browser back.

On October 7, 2025, Google released the Gemini 2.5 Computer Use model in preview. It is built on Gemini 2.5 Pro and is reached through a computer_use tool in the Gemini API, in Google AI Studio and Vertex AI. Google said it is aimed at web browsers, and that it was not yet aimed at desktop operating system control. Google also said versions of the model power Project Mariner.

On June 24, 2026, Google made computer use a built in tool in Gemini 3.5 Flash. Developers can use it through the Gemini API and the Gemini Enterprise Agent Platform. Google said the model can act in a browser, on mobile, and on a desktop. Two optional controls let a company require a person's confirmation, or stop a task if an indirect prompt injection is found.

Coding AI

On May 16, 2025, OpenAI launched Codex as a research preview. It is a cloud software engineering AI powered by codex-1, a version of OpenAI o3 trained for software work. Each task runs in its own sandbox with the repository loaded. Codex can write a feature, answer a question about the code, fix a bug, and propose a pull request. It started with ChatGPT Pro, Business, and Enterprise, with Plus to follow.

On October 6, 2025, OpenAI said Codex was generally available. New pieces that day were a Slack connection, a software kit, and admin tools. OpenAI said daily use had grown by more than 10 times since early August, and that GPT-5-Codex had served over 40 trillion tokens in the three weeks after it launched.

On May 19, 2025, GitHub introduced a Copilot coding agent. You assign a GitHub issue to Copilot, or ask it from VS Code. It works in a GitHub Actions environment, pushes commits to a draft pull request, and leaves a log of its steps. GitHub said those pull requests need a person's approval before CI workflows run. It was available to Copilot Enterprise and Copilot Pro+ customers. On September 25, 2025, GitHub said it was generally available for all paid Copilot subscribers.

On February 2, 2026, OpenAI released the Codex app for macOS. It is a place to run several AIs at once, in separate threads, with worktrees so they can share a repository without colliding. A March 4, 2026 update on that page says the app is also on Windows.

On February 5, 2026, OpenAI released GPT-5.3-Codex for paid ChatGPT plans, in the app, the command line, the IDE extension, and the web. OpenAI said the model is 25% faster for Codex users, and that a person can steer it while it works. In the table published with that post, GPT-5.3-Codex scored 56.8% on SWE-Bench Pro and 77.3% on Terminal-Bench 2.0.

Memory that carries forward

On February 13, 2024, OpenAI began testing memory in ChatGPT. The product can remember details across chats so a person does not have to repeat them. The person can ask what it remembers, tell it to forget, or turn memory off.

Updates on that same page say that on April 10, 2025, memory could also draw on past conversations, not only on a saved list. On June 3, 2025, a lighter memory update began rolling out to free users, with a longer view for Plus and Pro.

On June 4, 2026, OpenAI described a newer memory system it calls dreaming. It builds a fresh summary from many chats, and that summary can be read and edited. The rollout started with Plus and Pro users in the United States, with more countries and Free and Go plans to follow.

Sources

Opened and checked on November 11, 2026. Each date below is the date on the source.