Our Take
The real story is not the pricing change itself—it's that AI companies are discovering their unit economics don't work, and IPO pressures will force that discovery into public filings before anyone knows the answers.
Why it matters
As Anthropic, OpenAI, and others prepare S-1 filings, they must disclose cost and profitability risks in an industry where business models have shifted faster than the ability to measure them. Practitioners and investors are about to see how fragile the current AI cost model really is.
Do this week
Finance teams: audit your AI spending against per-token models now so you have real baseline costs before vendors finalize contract renewals this quarter.
Microsoft Moves Copilot to Per-Token Pricing
Microsoft announced pricing changes for GitHub Copilot, shifting away from flat-rate billing to a per-token model. The change was significant enough that internal teams at at least one company began referring to it as the "Tokenpocalypse." The move reflects broader pressure on AI companies to align pricing with actual usage and costs.
This shift arrives as major AI labs prepare for initial public offerings. Anthropic, OpenAI, and others will soon file S-1 registration statements with the SEC, forcing them to disclose profitability, cost structure, and material business risks (per the source discussion on upcoming IPO filings).
The timing mirrors a pattern already visible at large AI adopters. Uber, a heavy user of AI, burned through its planned AI budget faster than expected within weeks, then capped usage and limited employee access (company-reported). The speed of that reversal—from aggressive adoption to internal constraints—suggests cost discovery is happening faster than deployment strategies.
IPO Filings Will Force Transparency on Unsolved Economics
The core risk is not the price change itself. It is that AI companies have built pricing mechanisms without stable business models underneath them. The $20-per-month ChatGPT Plus subscription, for example, was never backed by formal strategy—it was a placeholder number (per the source). Even higher-tier pricing does not yet cover true operating costs.
The problem compounds at IPO time. Regulatory filing requires companies to disclose material risks and cost drivers. For AI labs, that means articulating how token demand, inference costs, and customer willingness to pay will eventually intersect. Those dynamics are still moving daily. As one source observer noted: "How do you even write these risks in, because they are evolving before your eyes?"
This is not a cyclical pricing correction. This is a forced reckoning with whether the current cost curve and revenue model can coexist. Investor-subsidized AI services will not survive public market pressure to show a path to profitability, especially when the path requires capabilities (cheaper inference, better efficiency) that do not yet exist at scale.
Contract and Spending Strategy Must Shift Now
For teams currently using AI APIs and services, the window to lock in grandfathered or flat-rate terms is closing. As vendors model for investor scrutiny and IPO filings, they will continue to shift pricing toward consumption models that reflect true marginal cost. That will likely include higher per-token rates, usage caps, and stricter SLAs.
The comparison to Uber is instructive but not reassuring. Uber eventually reached profitability, but only after years of unprofitability and only by fundamentally restructuring its cost base and supplier relationships (driver pay, geographic strategy, service mix). AI labs have fewer levers: they cannot easily reduce token generation costs without algorithmic breakthroughs, and they cannot arbitrarily cut customer compensation. The pressure will fall on pricing, usage limits, and feature tiering.
Practitioners should expect a period of rapid iteration on AI product pricing and availability before pricing settles. Companies already running material AI spend should document baseline costs under current pricing, plan for 30 to 50 percent price increases or usage restrictions within 12 months, and pressure vendors for multi-year contracts with caps if available. The alternative is reactive cost management when the market catches up to the actual economics all at once.