Meta Started Charging for Its AI Agents. The Framework Under Them Just Lost Its Creator to OpenAI

Meta's Business Agent stopped being free on August 1, 2026, and the price has held at $2.00 per million tokens since.

Mark Zuckerberg spent the first half of 2026 building a personal AI agent to help him run the company, according to reporting in The Wall Street Journal, while pushing every business on the platform toward an agent of its own. This page was re-checked against Meta's pricing documentation and investor releases on 1 November 2026.

That second agent, Meta AI's Business Agent, moved from a free test window to token-based billing on August 1: $2 per million tokens, roughly four to five cents per WhatsApp message. For an operator running a five-person shop through Instagram or WhatsApp, the tool pitched as a free sales rep now shows up as a monthly invoice.

What changed since the August review

Re-checked 1 November 2026: the price held at $2.00 per million tokens, and Meta lists one global rate with no volume tiers.

What actually changed since March

When this story first surfaced in March, the headline detail was novelty: an executive with a chatbot pipeline into internal dashboards. Seven months on, three things about it are concrete instead of hypothetical.

• Users: Meta AI reached 1.2 billion monthly active users in the first quarter of 2026, up from 1 billion in October 2025.
• Revenue: the advertising business posted $63.06 billion in the third quarter, about 97.6% of the $64.60 billion total.
• Spending: Meta's 2026 capital expenditure outlook is now $135 billion to $145 billion, narrowed in the third-quarter release from $130 billion to $145 billion, part of a build-out Zuckerberg has told investors could reach $600 billion by 2028.

None of that is about the CEO agent specifically. It is the balance sheet the CEO agent sits on top of.

Business Agent has a price, and a rival price card

The free test window ended on July 31. Metered pricing of $2.00 per million tokens took effect on August 1. Meta's developer documentation describes it as one per-token charge that covers both the AI processing and the message delivery. Meta charges for input and output tokens only, not for cache tokens, and volume tiers do not apply.

A typical message consumes 20,000 to 25,000 tokens, so most single-message interactions cost four to five cents. Meta says complex replies cost more than simple ones. A message that starts inside a free entry point window does not escape the charge either, because the documentation says Business Agent messages stay billable there.

Since October 1, Meta has also charged per message for what it calls service messages: any free-form reply that is not template-based and is not powered by Business Agent. That covers replies written by people and replies from third-party AI. Meta keeps a free monthly tier of service messages, and rates vary by market.

In its own side-by-side for Brazil, Meta estimates a third-party AI model at about 9 cents per message against 4 to 5 cents for Business Agent. Meta sells one side of that comparison, so treat it as the seller's estimate, not an independent test.

Earlier coverage described the agent as free to run inside the WhatsApp Business app and Meta's other business apps. Meta's current pricing page does not list any such exemption for Business Agent messages, so budget as if every agent reply is metered.

Who this helps

For a five-person shop that already lives on Instagram DMs, the pitch holds up: a sales agent trained on your product catalog and past posts, live inside the app your customers already use, for a few cents a message. HubSpot and comparable small-business marketing platforms charge a monthly seat fee whether or not a message goes out. Business Agent charges per use instead.

That math favors low-volume, high-margin sellers. It is a worse deal for a support-heavy business fielding hundreds of repeat questions daily, where a flat-rate tool can undercut token pricing fast.

If you are weighing a metered agent against a flat-rate one, answer a few Smart Match questions and get a shortlist matched to your volume and budget. The best AI chatbots and best AI agent builders roundups cover the flat-rate side, and the wider AI tools directory lists the rest.

Who this hurts

Customer-support software is the obvious target. Intercom and its peers built a business on sitting between a company and its Instagram or WhatsApp inbox, and Zendesk sells into the same inbox. Meta just became a direct competitor inside its own messaging apps, with no separate subscription required to sell against.

It also hurts the businesses that tested Business Agent for free through July, built workflows around it, and now absorb a per-message cost that was not part of the original pitch. Meta has not published an average per-business monthly cost under the new pricing, so anyone budgeting for it is estimating from message volume alone.

Anthropic's answer: Claude Duet

On 21 October 2026, the day of those earnings, Anthropic launched a preview of Claude Duet, pitched by CNBC as a rival to Muse. The app runs on Claude Opus 5.5, watches the apps you connect, and asks before it sends, pays or signs. Access starts with a waitlist for US subscribers on the Pro and Max plans.

The positioning is the useful part. CNBC notes that Meta has marketed its agent mainly to consumers. Anthropic says Duet is built for the workplace, so personal errands such as booking tickets are not as strong. It is early, the company warns, and can get a thread or an urgency call wrong.

For the small-business owner this guide has followed, the split is clear. Business Agent is a metered sales rep inside the customer inbox. Muse is the consumer hit: Zuckerberg told analysts that more than 10 million people in the US and Canada have set up one over the past month and a half. Duet is a chief of staff on the owner's side of the desk.

The framework Meta doesn't actually control

Meta's internal agents, including the CEO tool and an employee assistant, run on OpenClaw, an open-source framework that also powers Moltbook, the AI-agent social network the company acquired in March 2026 for an undisclosed sum. OpenClaw wraps several foundation models, including Claude, ChatGPT and Grok, so a task can go to whichever model handles it best. The OpenClaw playbook covers the framework in full.

In February, OpenAI hired OpenClaw's creator, Peter Steinberger, to build the next wave of personal agents, and moved OpenClaw itself into an independent foundation rather than under any single company's control. The software running the company's own leadership tooling is now stewarded by a structure that answers to neither Meta nor OpenAI, even though the latter just hired the person who built it.

OpenAI's own agent products sit on the other side of that line. Codex runs on its models, GPT-5.5 is the flagship behind them, and the OpenAI Agents API pricing guide shows how it meters computer use. Manus vs ChatGPT Agent compares the consumer versions.

The turn: this is less centralized than it looks

The obvious read is that Zuckerberg is building a surveillance layer for himself, one query away from every internal document at a company with roughly 70,000 employees after May's cuts. That is a fair worry, and it is not a new one.

A March 2026 incident, in which an internal agent auto-posted uncertain technical advice to a company forum, led another engineer to widen data permissions for nearly two hours. Meta classified it as a Sev-1, its second-highest internal severity tier, and said no user data was accessed by unauthorized parties.

The counter-argument is that querying dashboards faster is not the same as deciding better, and the company's own numbers back that up only partly. Meta's chief financial officer told investors output per engineer rose 30% since early 2025, with self-described power users up 80% year over year.

That is a productivity claim, not a governance one, and the two get conflated in most coverage of the CEO agent. Microsoft's Copilot pitches a similar executive-summary use case to enterprise customers, so he is not running this experiment alone. Coding agents such as Cursor and GitHub Copilot are where the engineer-productivity numbers come from, which is a different bet than a CEO dashboard.

Rivals frame the same bet differently, as Grok Bot shows in Grok Bot vs Meta Muse vs OpenAI Dots, and see the autonomy test with real money for what agents do when they hold a budget.

What to watch

Meta reported third quarter 2026 results after market close on October 21. It was the first report that fell entirely inside Business Agent's paid era, and the release gives no Business Agent revenue figure: total revenue was $64.60 billion, up 26% from a year earlier, and almost all of it was advertising. Until Meta names a Business Agent line item or a specific figure, how many free testers converted into paying accounts stays unknown.

Silence in a first paid quarter is not proof of failure, but it is also not proof of success, so a shop owner should budget from their own message volume and treat any vendor claim about average spend as unverified until a figure is published.

Watch too whether the OpenClaw foundation stays neutral or drifts toward OpenAI's own roadmap now that its creator works there. A forked, closed version of the framework would break the "nobody controls it" framing within a year. For wider context, see best agentic AI in 2026.

Zuckerberg wants an agent that lets him skip the org chart. Every business signing up for Business Agent is being asked to accept the same trade, minus the option to opt out of the bill.

Frequently asked questions

How much does Meta Business Agent cost now?

Meta charges one global rate of $2.00 per million tokens, in effect since August 1, 2026. A typical message uses 20,000 to 25,000 tokens, so about four to five cents per message. Meta's documentation says the charge covers both AI processing and message delivery, with no volume tiers.

Is Zuckerberg's personal AI agent the same product as Business Agent?

No. The CEO agent is an internal tool that surfaces Meta's own operational data for Zuckerberg, reported by The Wall Street Journal in March 2026. Business Agent is a customer-facing sales tool for outside businesses, and the two run on related but separate infrastructure.

What happened to OpenClaw, the framework behind Meta's internal agents?

OpenAI hired OpenClaw's creator, Peter Steinberger, in February 2026 to work on personal agents. OpenClaw itself moved into an independent foundation rather than staying under any single company, including Meta, whose own internal tools depend on it.

What was Meta's internal AI agent data exposure incident?

In March 2026, an internal agent auto-posted unreviewed technical advice to a company forum. Another engineer acted on it and widened data permissions for nearly two hours. Meta classified it as a Sev-1 incident and said no user data was accessed by unauthorized parties.

How many people use Meta AI as of 2026?

Meta AI reached 1.2 billion monthly active users in the first quarter of 2026, up from 1 billion in October 2025. It is built into Facebook, Instagram, WhatsApp and Messenger.

Covered in this guide

• Meta AI: Meta's free assistant inside Facebook, Instagram, WhatsApp and Messenger, powered by Muse Spark, with Meta One bundles for heavier image and video generation.
• ChatGPT: ChatGPT is OpenAI's AI assistant, with 1.2 billion weekly users by OpenAI's count, the GPT-6 model family and plans that run from a free tier to a premium Pro tier.
• Claude: Claude is Anthropic's AI assistant, with a free tier, paid personal plans, and Team and Enterprise seats. Opus access starts at the first paid plan.
• Claude Duet: Anthropic's research-preview agent on Claude Opus 5.5 watches your connected apps, messages you first, and waits for your OK before it sends, pays or signs.
• Claude Opus 5.5: Anthropic's September 2026 flagship LLM with a 1 million token context window, built for long-running agentic coding and computer use.
• Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
• GitHub Copilot: GitHub's AI coding assistant with agent mode and MCP support across 6 IDEs, plus a free plan with no credit card required.
• GPT-5.5: GPT-5.5 is OpenAI's April 2026 flagship model for agentic coding and knowledge work, with a 1.05M-token API context window and text and image input.
• Grok: Chatbot from xAI that searches the web and X live, now sold with Grok Bot and Cursor inside The X Subscription.
• Grok Bot: xAI's always-on AI agents that sign into your existing apps and finish multi-step jobs, included in xAI's joint Grok and Cursor bundle from the Premium tier up.
• Intercom: Customer service platform for SaaS teams: the Fin AI agent resolves most incoming queries automatically, and Intercom serves more than 30,000 organizations worldwide.
• Microsoft: Microsoft, founded 1975, is a public company (NASDAQ: MSFT) and the Redmond maker of Windows, Azure, GitHub and the Copilot assistants, with $331.8B of fiscal 2026 revenue.
• Microsoft Copilot: Your everyday AI companion powered by GPT-5.6 (preferred model July 2026), GPT-4o, and GPT-5 for productivity, creativity, and intelligent assistance
• Muse: Meta's AI agent runs tasks in its own cloud computer; open in the US, Canada, Japan and Singapore, with small-business skills for Shopify, QuickBooks and Meta ads in the US and Canada.
• OpenAI: OpenAI builds ChatGPT, the GPT-6 family (Astra, Sol, Luna), Codex and the OpenAI API. It reports 1.2B weekly ChatGPT users and an $852B valuation after its March 2026 round.
• Codex: Clones a GitHub repo into a sandbox, writes and tests code, then opens a pull request; OpenAI reported 20 million active users in August 2026.
• Zendesk: Omnichannel customer support platform with AI agents, ticket management, and 1000+ integrations for 100,000+ businesses.

Sources