OpenAI and Anthropic are two of the labs whose agent bets matter most to buyers, and for most of 2025 those bets looked like opposites. OpenAI showed Operator, a research preview of an agent that uses its own browser to do tasks for you. Anthropic published plumbing: the Model Context Protocol (MCP) for connecting models to tools (Anthropic, "Introducing the Model Context Protocol", November 2024), plus a computer use capability for Claude (Anthropic, "Introducing computer use", October 2024).

As of October 2026, that split has mostly closed. OpenAI puts its agents inside ChatGPT: ChatGPT Work handles longer, multi-step tasks and can drive its own cloud browser, and workspace agents let teams share agents in ChatGPT and Slack. Developers get the Agents SDK, the Responses API and a managed Agents API. Anthropic started with the plumbing and now ships its own agents on top: Claude itself (Cowork merged into it in September 2026), Claude in Chrome, Claude Code, and Claude Managed Agents for developers. Both labs also back the Linux Foundation's Agentic AI Foundation, which hosts MCP, OpenAI's AGENTS.md and, since October 22, 2026, Anthropic's Agent Skills (Anthropic, "Donating Agent Skills to the Agentic AI Foundation").

This guide maps both stacks layer by layer, lists what each lab has retired or merged since 2025, and gets practical about where to build or what to buy. Our verdict up front: the plumbing is shared now, so the choice that matters is which app and which hosted agent service you commit to.

Two abstract stacks of dark blocks, one orderly and one broken apart, each with an orange core, joined by a thin orange line

OpenAI vs Anthropic agents: from two bets to shared ground

In 2025 the two labs really did lead with different things. OpenAI first showed a browsing agent in its Operator research preview in January 2025, "a research preview of an agent that can use its own browser to perform tasks for you," available to Pro users in the U.S. (OpenAI, "Introducing Operator", January 23, 2025). Anthropic led with MCP and with computer use, which it launched in public beta in October 2024 (Anthropic, "Introducing computer use").

Those starting points still show, but the products have met in the middle. Operator no longer exists as a product; OpenAI points users to ChatGPT Work and workspace agents instead. By October 2026 each lab offers the same five layers:

The standards point is the easiest one to miss. The Linux Foundation formed the Agentic AI Foundation (AAIF) on December 9, 2025 "with founding contributions of leading technical projects including Anthropic's Model Context Protocol (MCP), Block's goose, and OpenAI's AGENTS.md," and its platinum members include both Anthropic and OpenAI (Linux Foundation). The rivals now back the same foundation that hosts the plumbing their agents run on. For where these two sit among the rest of the market, see our state of AI agents in mid-2026 overview.

In 2025 OpenAI led with Operator and Anthropic with MCP. By late 2026 both run agents inside their own apps and both build on MCP and Agent Skills under the Linux Foundation's Agentic AI Foundation (Linux Foundation, December 2025; Anthropic, October 2026).

What changed since 2025: retirements, merges and handovers

Some OpenAI vs Anthropic comparisons still online describe products that have since been retired. If you built on any of these, plan a migration. The table lists each change as of October 2026, from the vendors' own documentation.

Product Lab Status, October 2026 Where to go instead
Operator OpenAI Research preview from January 2025; folded into ChatGPT as ChatGPT agent on July 17, 2025 (OpenAI) ChatGPT Work and workspace agents
ChatGPT agent OpenAI "ChatGPT agent is no longer available" (OpenAI Help Center) ChatGPT Work, with its cloud browser
Assistants API OpenAI Removed from the API on August 26, 2026 (OpenAI deprecations) Responses API and Conversations API
computer-use-preview model OpenAI Shut down on July 23, 2026; it was the API model behind Operator (OpenAI deprecations) gpt-5.6-terra, the substitute OpenAI's deprecations page names; the current computer use guide, which defaults to GPT-6 Astra
Agent Builder, Evals platform, reusable prompts OpenAI Shutting down on November 30, 2026; Evals goes read-only on October 31, 2026 (OpenAI deprecations; AgentKit update) Agents SDK for code, workspace agents for natural language; ChatKit remains available
Claude Cowork Anthropic Merged into Claude on September 16, 2026, rolling out on Pro and Max plans first (Anthropic) Claude
"Building effective agents" tooling advice Anthropic The December 2024 essay now notes that much of the tooling it describes has changed (Anthropic) Claude Managed Agents

OpenAI's list is longer because it has shipped, renamed and retired more agent products since 2025. Its migration guide offers two paths off Agent Builder: the Agents SDK, "best for building agents through code," and workspace agents, "best for building agents through natural language and sharing them with teams" (OpenAI). For the full picture on that shutdown, see our guide to OpenAI Agent Builder alternatives.

The other change is who owns the standards. MCP now runs as "Model Context Protocol a Series of LF Projects, LLC," with code and specification contributions under the Apache 2.0 license (MCP governance), and the current spec version is 2026-07-28 (MCP versioning).

On October 22, 2026 Anthropic donated Agent Skills to the AAIF as well. A skill is "a folder containing a SKILL.md file that teaches an agent a specific task or domain," and Anthropic reports that "more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Agent Skills" since it published the standard in December 2025. That count is Anthropic's own, with no independent tally behind it. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals (Anthropic).

OpenAI's agent strategy

OpenAI's agent strategy starts from ChatGPT. Its in-app agent is ChatGPT Work: "Work is an agent designed for longer, multi-step work and finished deliverables," while "Codex remains dedicated to software development and technical work" (OpenAI Help Center). OpenAI lists ChatGPT Work on all plans in the macOS and Windows desktop apps, and on Plus, Pro, Business, Enterprise and Edu on web and mobile (OpenAI). For browser tasks, "cloud browser gives ChatGPT Work its own browser on a separate computer in the cloud," on paid ChatGPT plans in supported regions, excluding Free and Go (OpenAI Help Center). Our explainer is ChatGPT an AI agent covers how ChatGPT Work fits the definition.

For teams, OpenAI has workspace agents, which it calls "an evolution of GPTs." They are powered by Codex, generally available in ChatGPT Business, Enterprise and Edu, and usable in both ChatGPT and Slack, and "you can require the agent to ask for permission before moving forward" (OpenAI, "Introducing workspace agents in ChatGPT", April 22, 2026).

Starting from ChatGPT's reach

OpenAI still tends to ship the whole experience first. When a capability is ready, it usually arrives inside ChatGPT, tuned and rate-managed by OpenAI, so a non-technical user can get value without wiring anything. The road from Operator to ChatGPT agent to ChatGPT Work shows the pattern: three names since January 2025, each folded into the main app. That suits buyers who want results on day one. The cost is that the product you learned can be renamed or retired under you.

Tooling and the developer surface

For developers, OpenAI's agents guide now offers four starting points (OpenAI, Agents guide):

Function calling, the Responses API, the Agents SDK and a large library of community wrappers mean you rarely start from scratch. OpenAI's developer tools are also more open than they were. The API can give models "new capabilities using remote MCP servers or Secure MCP Tunnel" (OpenAI, MCP guide), and its skills "are compatible with the open Agent Skills standard" (OpenAI, Skills guide). The models behind all of this are GPT-6 Astra, the flagship "for complex reasoning and coding," GPT-6.1 Sol to balance intelligence and cost, and GPT-6 Luna for cost-sensitive, high-volume work (OpenAI, Models). For one concrete thread, see our take on OpenAI's Codex agent implications.

OpenAI's agent strategy runs through ChatGPT: ChatGPT Work for longer tasks, with its own cloud browser, and workspace agents for teams (OpenAI Help Center). For developers it offers a managed Agents API (OpenAI, Agents guide) and an Agents SDK that also runs other providers' models (OpenAI Agents SDK README).

Anthropic's agent strategy

Anthropic's agent strategy started with plumbing. Its most distinctive early move was MCP, which it open-sourced in November 2024 as "a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments" (Anthropic, "Introducing the Model Context Protocol"). It paired that with computer use, a documented way for developers to have Claude work a computer as a person does: looking at a screen, moving a cursor, clicking buttons and typing text (Anthropic, "Introducing computer use", October 2024). Anthropic called computer use "public beta" and "still experimental" at launch. The computer use tool is now generally available on the Claude API with the computer_toolset_20260801 release (Claude docs).

Since then Anthropic has built finished agents on top of that plumbing:

On the model side, Anthropic's models page lists Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. It recommends that you "start with Claude Opus 5.5 for most workloads" and use Fable 5.1 "for demanding reasoning and long-horizon agentic work" (Claude docs, Models overview). Opus 5.5 shipped on September 22, 2026, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7 (Anthropic newsroom). Our earlier piece on Anthropic Claude agent capabilities covers how Claude works as an agent backbone.

From protocol to products

Anthropic still tends to ship a capability and a standard first, then a product on top. MCP spread well beyond Anthropic. In December 2025 Anthropic reported that "MCP has been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot, Visual Studio Code, and other popular AI products" (Anthropic). An open protocol that competitors adopt becomes industry infrastructure, and MCP now literally is: it sits under the Linux Foundation. Agent Skills has followed the same path, about ten months after Anthropic published it as an open standard.

Safety as a design input

Anthropic's best-known agent guidance is its December 2024 essay "Building effective agents." It found that "the most successful implementations weren't using complex frameworks or specialized libraries. Instead, they were building with simple, composable patterns," and it recommends "finding the simplest solution possible, and only increasing complexity when needed" (Anthropic, "Building effective agents"). It also warns that autonomy brings "higher costs, and the potential for compounding errors," and advises extensive testing in sandboxed environments with guardrails.

The essay now carries a note that much of the tooling it describes has changed since December 2024, and it sends readers to Claude Managed Agents for Anthropic's current approach. The design advice still holds; the recommended tools have moved.

Anthropic's agent strategy started with open plumbing, MCP in 2024 and Agent Skills in 2025, and has since added its own agents: Claude, Claude in Chrome, Claude Code and Claude Managed Agents (Claude docs). It has donated both standards to the Linux Foundation's Agentic AI Foundation (Anthropic, October 22, 2026).

The two agent stacks side by side

The strategies now rhyme on capability and on standards. The split shows up in where each lab starts and how much of the experience it wants to host for you. Here is each stack in one view, built from both labs' own documentation as of October 2026 and sourced in the sections above. Treat it as a snapshot: as the retirements table shows, OpenAI in particular renames and retires products quickly.

Layer OpenAI (October 2026) Anthropic (October 2026)
Agent inside the app ChatGPT Work, with a cloud browser on paid plans except Free and Go Claude (Cowork merged in on September 16, 2026, rolling out on Pro and Max); Claude in Chrome, generally available on every paid plan
Shared team agents Workspace agents in ChatGPT Business, Enterprise and Edu, usable in ChatGPT and Slack None named in Anthropic's docs cited above
Hosted agent for developers Agents API: your agent runs on the framework behind Codex, with OpenAI managing the infrastructure Claude Managed Agents (beta), for long-running and asynchronous work
SDK you run yourself Agents SDK, which supports 100+ other LLMs Claude Agent SDK: Claude Code as a library, in Python and TypeScript
Raw API Responses API (the Assistants API was removed on August 26, 2026) Messages API
Coding agent Codex Claude Code
Computer use on the API Computer use guide, running on GPT-6 Astra Computer use tool, generally available on the Claude API
Open standards Co-founded the AAIF and contributed AGENTS.md; the API supports MCP and Agent Skills Co-founded the AAIF; created MCP (donated December 2025) and Agent Skills (donated October 22, 2026)
Retired, merged or retiring ChatGPT agent, the Assistants API and computer-use-preview are gone; Agent Builder, the Evals platform and reusable prompts follow on November 30, 2026 Claude Cowork, merged into Claude on September 16, 2026
Current models GPT-6 Astra (flagship), GPT-6.1 Sol, GPT-6 Luna Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5
Safety emphasis Research previews and staged rollouts; a review model screens sign-in requests in the cloud browser Safety foregrounded in published agent design guidance; a classifier checks each Claude in Chrome action

Autonomy: how far to let the agent go

Both labs let an agent chain steps, and both now let you loosen the leash. ChatGPT's cloud browser offers three permission modes, "Always ask," "Auto approve" and "Always allow," and OpenAI marks the last one "not recommended" (OpenAI Help Center). Claude in Chrome can now act in the browser without approval for every step. The OpenAI Agents SDK lists guardrails and human in the loop among its core concepts, while Anthropic's published guidance still leans toward agents that are simple and composable. The rhetoric about how much rope to give differs more than the controls do.

Openness: who owns the plumbing

This used to be the cleanest divide; it is now the clearest convergence. Both labs put core pieces in the open. Anthropic donated MCP, and in October 2026 Agent Skills, to the Agentic AI Foundation. OpenAI contributed AGENTS.md, which the foundation describes as "a simple, open format for guiding coding agents" (AAIF). The foundation's project list also includes Block's goose and Agent Router.

So where does lock-in sit? In the apps, ChatGPT and Claude, and in the hosted agent services. If you care about portability, that is where to look before you commit. For how MCP compares with A2A and other agent protocols, see our breakdown of the AI agent protocol wars.

Safety: two browser safeguards

Safety framing differs in emphasis, and both labs now ship named safeguards on their browser agents. In ChatGPT Work's cloud browser, "before presenting a sign-in request, an additional review model checks the request and where your information will be entered for signs of phishing or deception" (OpenAI Help Center). In Claude in Chrome, "a safety classifier validates each action before it's performed to ensure it's safe and matches your request" (Anthropic). OpenAI also tends to launch agent features as research previews first, as it did with Operator and workspace agents. Neither safeguard replaces your own review on actions that move money or data.

Developer and enterprise approach

For developers and enterprises, the choice rarely comes down to a single benchmark. It comes down to fit: how the agent surface matches your stack, your risk posture and your tolerance for vendor dependence. Both labs offer enterprise terms, API access and documented agent capabilities, so the useful question is whose shape fits your team.

OpenAI's pitch to a team is reach and polish. Workspace agents run inside ChatGPT and Slack, and the developer stack runs from a fully managed Agents API down to raw Responses calls. The cost is pull toward ChatGPT-specific features, and OpenAI has shown it will retire them: developers on the Assistants API got one year's notice before its removal, and Agent Builder users got about six months (OpenAI deprecations).

Anthropic's pitch leans on developer and coding tools. Claude Code and the Claude Agent SDK share the same tools and agent loop, and Claude Managed Agents lets Anthropic host the agent when you would rather not run the infrastructure yourself. The trade is that Managed Agents is still in beta, and the Cowork merge into Claude started on Pro and Max, with other plans to follow.

The enterprise risk calculus

Enterprises weigh agents differently than hobbyists. They ask about auditability, data handling and what happens if a vendor changes terms or direction. In 2026 that last question has concrete answers: OpenAI's deprecations page gives dated shutdowns with named replacements, and Anthropic's own agents essay now points readers to its newer product. Anthropic's safety-forward framing reads well to a risk committee. OpenAI's maturity and reach read well to a delivery team under deadline. Both arguments are legitimate, and the right answer is org-specific.

Where the two converge

The labs have landed on the same ideas. Computer use appeared on both sides, and so did support for the same open standards, MCP and Agent Skills. Structured tool use is universal now: OpenAI's function calling guide describes it as a way for models "to interface with external systems," and MCP does the same job for any model that supports it. Even the developer shapes match, with a hosted agent service, an SDK, a raw API and a coding agent on each side.

Spend your effort on scoping the task clearly and wiring its tools well. That work carries over whichever lab supplies the model.

What it means for builders and buyers

For builders and buyers, the practical takeaway is to design for optionality. The two labs are converging on capability and on standards. They differ in which apps they want you to live in, so your biggest lever is keeping your workflow loosely coupled to any one provider, which lets you ride improvements from both without a rebuild.

If you build, lean on the open standards. Wiring tools through MCP keeps your integrations portable, and writing reusable instructions as Agent Skills does the same for know-how. OpenAI's skills guide describes a skill as "a directory of files with a SKILL.md manifest" (OpenAI), the same format Anthropic donated to the AAIF. If you buy, ask the vendor a simple question: can you switch the underlying model without redoing the work? For how this plays out at the workspace level, see Claude vs ChatGPT workspace agents.

Using both labs together

You can use both, and the tooling now expects it. Three things make it practical:

  1. One SDK, several providers. The OpenAI Agents SDK supports "100+ other LLMs" besides OpenAI's own APIs (GitHub), so one agent loop can call models from more than one lab.
  2. One tool server, both apps. Both labs speak MCP: OpenAI's API connects to remote MCP servers, and Anthropic lists ChatGPT among MCP's adopters. A tool you expose once over MCP can serve both.
  3. One skill format. OpenAI's skills follow the open Agent Skills standard, and Anthropic counts ChatGPT among the platforms that have integrated it, so a SKILL.md folder written once can travel.

The catch is the app layer. ChatGPT Work and workspace agents come with ChatGPT plans, and Claude in Chrome with paid Claude plans, so using both apps means two subscriptions. Our ChatGPT pricing and Claude pricing explainers cover what each plan includes.

Common mistakes to avoid

Three mistakes to avoid:

  1. Hard-coding against one lab's app-specific agent feature because it is convenient today, then paying for it at migration time. The Assistants API and Agent Builder are this year's examples.
  2. Starting on a product that is already on a deprecation list. Check OpenAI's deprecations page before you build; Agent Builder's shutdown was announced on June 3, 2026, almost six months before the product goes dark.
  3. Treating the model choice as the whole decision and skipping the scoping and tool wiring the agent depends on. Our primer on what an AI agent is separates the model from the agent around it.

How to choose for your use case

Match the starting point to where your team already works. On a ChatGPT Business, Enterprise or Edu plan, workspace agents are the built-in way to share an agent across the team, while developers who want the agent loop inside their own application use the OpenAI Agents SDK. A team that builds in code first has a parallel path at Anthropic: Claude Code for coding work, the Claude Agent SDK for the same agent loop as a library, and Claude Managed Agents, still in beta, when you want Anthropic to host long-running jobs.

For browser work on internal tools, compare the two browser agents directly: ChatGPT Work's cloud browser runs on a separate computer in the cloud, while Claude in Chrome works in your own browser, which Anthropic pitches at internal dashboards, legacy systems and vendor portals. To limit your dependence on either lab, keep your tools in MCP and your instructions in Agent Skills.

Where a platform like Gravity fits

Gravity sits one level above the lab choice: you type the task and Gravity matches an expert-built agent that runs it. Recurring automations can run on a schedule, so you set a task once rather than rebuilding it when a lab changes its product line.

Gravity is in public beta, and your first 3 agent runs are free. The Free plan costs $0 with no card, Mini Pilot is $15 a month, and Auto Pilot is $20 a month. If you run out, you can buy more usage; the pricing page has the details.

If you are comparing directly, we keep two even-handed breakdowns: Gravity vs Claude and Gravity vs ChatGPT.

Frequently asked questions

What is the core difference between OpenAI and Anthropic on agents?

Less than it was in 2025. Both labs now ship finished agents inside their apps: OpenAI has ChatGPT Work and workspace agents, and Anthropic has Claude, Claude in Chrome and Claude Code. Developers get a hosted agent service from each (OpenAI's Agents API, Anthropic's Claude Managed Agents) plus an SDK. Both support MCP and Agent Skills, which now sit in the Linux Foundation's Agentic AI Foundation. The difference is where each started: OpenAI from ChatGPT's consumer reach, Anthropic from developer and coding tools.

What makes Anthropic different from OpenAI?

On agents, the two labs have converged more than their reputations suggest. Anthropic started with developer plumbing: it created MCP in 2024 and Agent Skills in 2025, and has donated both to the Linux Foundation's Agentic AI Foundation. OpenAI started with ChatGPT and now runs agents there through ChatGPT Work and workspace agents. Today both ship app agents, hosted agent services and SDKs. Anthropic leans on coding and developer tools such as Claude Code; OpenAI leans on ChatGPT's reach.

Is the Model Context Protocol an OpenAI or Anthropic project?

Anthropic created MCP and open-sourced it in November 2024. In December 2025 it donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation that Anthropic co-founded with Block and OpenAI. MCP is now run as an LF project under the Apache 2.0 license, and ChatGPT, Cursor, Gemini, Microsoft Copilot and VS Code have adopted it. For how an MCP server differs from an agent, see our guide to AI agents vs MCP servers.

Should I build my agent on OpenAI or Anthropic in 2026?

It depends on where you want the agent to run. If your team already works in ChatGPT, workspace agents and the OpenAI Agents SDK are the shortest path; if you build in code, Claude Code, the Claude Agent SDK and Claude Managed Agents are Anthropic's equivalents. Both support MCP for tools and Agent Skills for reusable instructions, so building on those keeps your work portable. The OpenAI Agents SDK can also run models from other providers.

What is computer use and does it favor one lab?

Computer use lets a model operate a screen the way a person does, moving a cursor, clicking and typing, so it can act in software that has no API. Both labs offer it. On the API, OpenAI's computer use guide runs on GPT-6 Astra and Anthropic's computer use tool is generally available on the Claude API. In the apps, ChatGPT Work can run its own cloud browser and Claude in Chrome works in your own browser on paid Claude plans. Neither lab has the field to itself.

What replaced OpenAI Operator?

Operator launched as a research preview in January 2025. In July 2025 OpenAI folded it into ChatGPT as ChatGPT agent, and ChatGPT agent has since been retired too. OpenAI now points users to ChatGPT Work, which can use its own cloud browser on paid plans except Free and Go, and to workspace agents for teams. The API model behind Operator, computer-use-preview, was shut down on July 23, 2026.

Does choosing a lab lock me into that vendor for agents?

It can, if you build directly against one lab's app-specific agent features. The plumbing is now shared: MCP and Agent Skills work with both labs and are hosted by the Linux Foundation's Agentic AI Foundation, and the OpenAI Agents SDK supports other providers' models. Lock-in now sits mostly in the apps and hosted agent services, so keep your tools and instructions in those open formats.

Sources

OpenAI

Anthropic

Standards bodies