A year ago, “AI agent” sounded like a lab demo. In September 2026, the four biggest companies in the industry all sell one, and each has placed it exactly where its users already were: OpenAI inside ChatGPT, Google inside Gmail, and Microsoft inside Word and Excel. Anthropic took the opposite route: it started in programmers' terminals and spread from there to everything else.

That is the key to understanding the landscape: rather than competing to have the smartest agent, they are competing to become a tool you can't do your job without. And they are moving so fast that in September alone, OpenAI retired its ChatGPT agent and Anthropic stopped calling its agent Cowork and renamed it simply Claude. This article covers what each company offers as of September 27, 2026, with prices and links to their official pages.

First, what counts as an agent

Before comparing, it helps to know what exactly we are comparing. Anthropic puts it simply: in a workflow, the model follows a path someone wrote in advance, whereas in an agent, the model itself decides which steps to take and which tools to use along the way (Schluntz & Zhang, 2024). Not everything companies sell as an agent is actually an agent by the strict definition. If you want the full definition, read our article what are AI agents.

OpenAI: from chatbot to coworker

OpenAI starts with a clear advantage: for many people, ChatGPT is simply synonymous with artificial intelligence. Its bet is that the same chat will stop merely answering questions and start doing real work. Its first attempt in that direction, ChatGPT agent, has already been retired: its help page now only says the product is unavailable and points to its successor (OpenAI Help Center, n.d.-a).

That successor is ChatGPT Work, an agent designed for long, multi-step assignments that delivers finished work instead of an answer. It runs on the web, on mobile, and on desktop with the paid plans that include it, and it sits alongside Codex, which remains OpenAI's specialized coding agent (OpenAI Help Center, n.d.-b). I compare Codex with its rivals in Claude Code vs Codex vs Cursor.

For those who want to build their own agents, OpenAI offers two routes: the Agents SDK, for Python and TypeScript, in which the developers themselves control the agent loop, and the Agents API, in which OpenAI takes care of running it (OpenAI, n.d.-a, n.d.-b). Anyone using Agent Builder, its visual tool, should start looking for an alternative, because OpenAI will shut it down on November 30, 2026 (OpenAI, n.d.-d). For businesses, OpenAI is also testing workspace agents, which an entire team shares and which can be scheduled to repeat tasks; for now they are only available on the Business, Enterprise, and Edu plans (OpenAI, n.d.-c).

Google: the agent that lives in your inbox (except in Europe)

Google holds a card no one else has: your email, your documents, and your calendar are already on its servers. In May 2026 it introduced Gemini Spark, a personal agent that works around the clock connected to Gmail, Docs, and Slides, designed to anticipate your needs and handle tasks without you having to ask every time (Google, 2026). In the United States it is included in the Google AI Pro subscription, at $19.99/month, and in Ultra, from $99.99/month (Google, n.d.-c).

The problem, if you are reading this from Europe, is that you can't use it: Spark is only available to people over 18 with a personal account, and it is not offered in the European Economic Area, the United Kingdom, or Switzerland (Google, n.d.-b). Of the four agents in this article, it is the one the fewest people can try.

Where Google is generous is with developers. Its Agent Development Kit (ADK) is open source, works with Python, TypeScript, Go, Java, and Kotlin, and is compatible with almost any generative AI model, so it doesn't force you to use Gemini (Google, n.d.-a). For businesses, the Gemini Enterprise app costs from $21 per user per month (Google Cloud, n.d.), and underneath it sits the Gemini Enterprise Agent Platform, which since April 2026 brings together everything previously known as Vertex AI (Google Cloud, 2026a). Its no-code agent builder, however, is still in preview (Google Cloud, 2026b).

Microsoft: the agent inside Office (with a borrowed brain)

Microsoft doesn't need you to open anything new: its agent lives in the program you already have open. Since April 2026, Copilot has worked as an agent inside Word, Excel, and PowerPoint, and it is turned on by default in Microsoft 365 Copilot and Microsoft 365 Premium (Microsoft, 2026b). In the Edge browser, Microsoft is testing Browse with Copilot, a feature that clicks, types, and navigates for you, although it is still experimental and only available in the United States (Microsoft Support, n.d.).

This is where the most curious detail of the whole landscape shows up. Copilot Cowork, Microsoft's agent for long assignments, available worldwide since June 2026, runs under the hood on models from Anthropic, one of its competitors (Microsoft, 2026c). Microsoft provides the house, meaning Office and your files, and another company provides the brain. It sums up how this market works: what is being sold is not so much the model as the place where it is used.

To build your own agents, Microsoft offers two options. Copilot Studio lets you do it without coding, with the Microsoft 365 Copilot license at $30 per user per month plus usage-based credits (Microsoft, n.d.). Agent Framework, aimed at developers, has combined AutoGen and Semantic Kernel since April 2026 and works with .NET and Python (Microsoft, 2026a). On top of that comes Foundry Agent Service, which can even run agents built with the OpenAI and Anthropic kits (Microsoft Learn, 2026). And because Microsoft knows companies will end up with dozens of scattered agents, it sells Agent 365, at $15 per user per month, to monitor and manage them all from one place (Microsoft, 2026d).

Anthropic: from the programmer's terminal to the office

Anthropic took the opposite path from the others. Instead of starting with the general public, it started with programmers, with Claude Code, an agent that works in the terminal, in the editor, on the desktop, and on the web (Anthropic, n.d.-b).

It then brought that same way of working to the office with Cowork, which extended Claude Code's capabilities to tasks that have nothing to do with programming (Claude Help Center, n.d.). On September 16, 2026, Anthropic folded it into its main app: the chat and the agent became a single product, simply called Claude, which is rolling out first to the Pro and Max plans (Anthropic, 2026).

For developers there is the Agent SDK, which provides the same tools, the same agent loop, and the same context management that Claude Code uses, for building your own agents in Python or TypeScript (Anthropic, n.d.-a). Managed Agents, still in beta, runs long-running agents on Anthropic's infrastructure (Anthropic, n.d.-c). As for pricing, the free plan does not include Claude Code; Pro costs $20/month, Max starts at $100/month, and Team at $25 per user per month (Anthropic, n.d.-d).

All four at a glance

If each strategy had to be summed up in one line: OpenAI wants ChatGPT to do your work, Google wants its agent to live in your inbox, Microsoft wants it inside Office, and Anthropic wants what worked for programmers to work for everyone. The table puts the products side by side:

One detail unites all four: their kits for building agents are open source (OpenAI, n.d.-b; Google, n.d.-a; Microsoft, 2026a; Anthropic, n.d.-a). The competition is in the finished products, not in the parts.

So which one is right for you?

The short answer: the one that is already where you work. If your company uses Microsoft 365, Copilot is the natural choice, because it is already inside Word, Excel, and PowerPoint and nobody has to switch tools. If it uses Google Workspace, Gemini Enterprise. And if you already pay for ChatGPT or Claude, the sensible thing is to try the agent you already have included before paying for another one.

If you live in Europe and want a personal agent, rule out Gemini Spark for now, since it isn't available in the EU (Google, n.d.-b). If you code and want to build your own agents, any of the four kits will do, although ADK is the most open to models from other companies (Google, n.d.-a). And if you don't want to send your data to any of these companies, there is an alternative: local AI agents, which run on your own computer.

One last piece of advice: these products change so quickly that what has one name today may have another next month, as happened to ChatGPT agent and Cowork within a matter of weeks. Before signing up for an annual plan, check the product's official page.

Frequently asked questions

What are the 5 types of AI agents?

The classic classification distinguishes five: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. I explain them with examples in what are AI agents.

Who are the big 4 AI agents?

Those from the four biggest companies in the industry: ChatGPT Work and Codex, from OpenAI; Gemini Spark, from Google; Copilot, from Microsoft; and Claude and Claude Code, from Anthropic (OpenAI Help Center, n.d.-b; Google, n.d.-b; Microsoft, 2026b; Anthropic, 2026).

What are the top 5 AI agents?

It depends on the use. For coding, Claude Code, Codex, and Cursor stand out; for office work, Copilot in Microsoft 365, ChatGPT Work, and Claude. None of them is the best at everything.

What is Anthropic AI and OpenAI?

Two American artificial intelligence research companies. OpenAI develops ChatGPT and Codex; Anthropic develops Claude and Claude Code. They compete with each other, although their models sometimes show up in third-party products: Microsoft's Copilot Cowork runs on Anthropic models (Microsoft, 2026c).

Does Google have an AI agent?

Yes. For individuals it offers Gemini Spark, with AI Pro or Ultra, although it isn't available in the EU; for businesses, Gemini Enterprise; and for developers, ADK (Google, n.d.-a, n.d.-b; Google Cloud, n.d.).

What is the difference between Microsoft Copilot and Copilot Studio?

Copilot is the assistant, which already works as an agent inside Word, Excel, and PowerPoint. Copilot Studio is the tool for building your own agents, and you pay for it with usage-based credits (Microsoft, 2026b; Microsoft, n.d.).

Keep reading

• →What are AI agents? Explained with examples
• →AI coding agents: Claude Code vs Codex vs Cursor
• →Local AI agents: how to run them on your own computer
• →AI agents built a secret message board — then broke into Hugging Face
• →All blog articles