Elena Voss$30. That's the monthly per-seat price Microsoft set for Copilot for Microsoft 365 in July 2023. It seemed reasonable then — a premium for AI-augmented productivity, positioned below the cost of an incremental hire, above the cost of the Office suite itself. Three years and billions of dollars in infrastructure investment later, a Microsoft senior executive called the number a "pretty major barrier" to sales.
Bloomberg reported that Judson Althoff, Microsoft's chief commercial officer, told internal teams that Copilot sales had hit "some pretty major barriers." This follows a pattern: by February 2024, early testers were already questioning the price-to-value ratio. By June 2025, interviews with 24 Microsoft customers and salespeople showed persistent friction. Now the company has responded by bundling Copilot into a new $99/user E7 tier — diluting the price signal by embedding the AI tool inside a larger package rather than defending the standalone price.
The Comparison
Cursor crossed $1 billion in ARR in November 2025. By February 2026, it had doubled. The company has 300 employees. The growth rate is roughly 10x what Microsoft's AI productivity tools are managing — and the pricing model is fundamentally different. Cursor's base plans are per seat, but the AI usage layer is consumption-based: each plan includes a credit pool tied to token consumption, and heavy users burn through it faster and pay more. Tab completions are unlimited. What you're actually paying for scales with how much you ask the agent to do.
The distinction matters because AI coding tools don't generate equal value per seat. Usage is bursty. The best developers run 10 to 15 parallel sessions. Junior developers may not use the tool at all. A per-seat model charges the heavy user and the non-user the same $30 — which means either the heavy user is getting a bargain or the non-user is subsidizing them. At enterprise scale, the subsidy math breaks down. The CFO sees $30 times 10,000 seats and asks what the company got for $3.6 million.
The Unbundling
Anthropic made a different adjustment. Claude subscriptions will no longer cover usage on third-party tools — the OpenClaws, the Cursors, the MCP-connected agents that route work through Claude's API. The subscription that used to mean "use Claude however you want" now means "use Claude in our apps." Everything else costs extra.
This is unbundling — separating the subscription from the ecosystem that made it valuable. It follows a pattern: in January, Anthropic added safeguards to prevent third-party apps from freeloading on Claude subscriptions. In March, they launched Claude Marketplace to give third-party integrations a revenue path. Now they've formalized the split.
The timing is not coincidental. The same agent infrastructure that makes Claude useful through OpenClaw and Cursor also makes the flat subscription economically indefensible. When one subscriber generates 100x more API calls than another because they're routing agent loops through the subscription, the flat-rate model leaks money in the direction of the heaviest users.
The Unit of Value
Three pricing adjustments in 48 hours. Each is a different answer to the same question: what's the right unit of value for AI software?
Microsoft's answer: bundle it and stop defending the standalone price. Anthropic's answer: unbundle it and charge for what's actually being consumed. Cursor's answer: let the base seat be a floor, then bill the AI usage layer by consumption — so the developer running 15 agent sessions pays more than the one running one.
The per-seat model worked for Office because every seat used Word roughly the same amount. It worked for Slack because every seat sent messages. It doesn't work for AI coding tools because the distribution of value per seat is wildly uneven. The developer running overnight research loops through Claude Code generates orders of magnitude more value — and more cost — than the developer who asks Copilot to autocomplete a function signature once a week.
$30/seat was a guess about how AI tools would be used. Three years in, the usage patterns answered: not like Office. Not like Slack. Like infrastructure — bursty, concentrated, and worth far more to the heavy user than the average price implies.
Cursor's growth suggests the market already knows this. Usage-based billing aligns the cost with the value: you pay in proportion to what the agent does. The one-person company running 15 agent sessions simultaneously pays more than the team that runs one — and both are paying in proportion to what they receive. The growth to $2B ARR in under two years, without the distribution advantage of Windows and Office, is the market's verdict on which model fits.
The Number, Recontextualized
$30 was the right price for the wrong model. It assumed per-seat uniformity in a domain defined by per-user variance. The AI tools that will win on pricing aren't the ones that charge less per seat — they're the ones that make the consumption layer visible and bill it separately. Cursor keeps a seat floor but bills the AI usage layer by actual token consumption. Open-source alternatives charge nothing for the weights and everything for the guarantees. Anthropic is unbundling so it can price the API by consumption.
The $30 that Microsoft set in 2023 is a fossil. Not because it was wrong at the time — but because AI tools turned out to be infrastructure, not features, and infrastructure has never been priced per seat.
More on Microsoft, Anthropic, Cursor, and AI pricing. Explore entity coverage via the Pulse API.