By Grego
I’ve been a CIO for several decades. I’ve watched pricing cycles come and go — shared time on mainframes, SaaS, compute in the cloud. I know how to spot a loss-leader when I see one, and I know what happens when it ends.
What’s happening right now with AI pricing is the most aggressive loss-leader program I’ve seen in the history of enterprise software. And most CTOs and CFOs in Iberoamerica still haven’t done the math.
Let me do it for you.
The gap between what you pay and what it actually costs
Claude Pro: $20/month per seat. For that price, your team gets access to Sonnet 4.6, Opus 4.6, web search, code execution, and document analysis. Seems reasonable.
Now look at API rates. Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens. Opus 4.6 runs $5 input and $25 output. A knowledge worker using Claude a few hours a day — uploading documents, drafting analyses, iterating on code — can easily consume several million tokens per week. At API rates, that same workload costs between $200 and $400 per month per seat. With a Pro subscription, the company pays $20.
That’s not a discount. It’s a subsidy. A widely cited analysis found that Anthropic users were consuming about $8 in compute for every $1 of subscription revenue.
Anthropic isn’t alone. Microsoft reported losses of more than $20 per user per month on GitHub Copilot. For heavy users, the compute cost reached $80 per month on a $10 subscription.
The math is the same across all major providers. OpenAI’s own VP of Product, Nick Turley, described their subscription prices as something they “stumbled into” and raised the possibility of eliminating unlimited plans, comparing them to “unlimited electricity.”
Why they’re doing it
This isn’t a calculation error. It’s a strategy.
The pattern is identical across all providers: price for adoption, not for economics. Build dependency. Make AI a structural part of every team’s daily workflow. Worry about the bill later.
For companies, “later” is arriving. The question for your organization isn’t whether you got a good deal on team pricing. The question is how long that deal lasts — and whether your budget is ready for what comes next.
Agentic use broke the underlying economics
What’s accelerating the timeline is agentic AI.
When AI was a chatbot, token consumption was relatively predictable. One conversation runs through a few thousand tokens. Heavy use might reach tens of thousands. That was manageable at subsidized flat rates.
The agentic shift changes everything. Claude Code sessions run autonomously for extended periods, consuming tokens at a pace that dwarfs conversational use. Users reported exhausting 5-hour rate limit windows in less than 90 minutes.
GitHub announced that Copilot is migrating to usage-based billing on June 1, 2026, specifically because the flat-rate model collapsed under agentic workloads. That’s not an isolated price change. It’s a canary in the coal mine.
Sam Altman said publicly that OpenAI needs to become “an AI inference company” — an acknowledgment that agentic use requires a fundamentally different economic model.
For engineering leaders: a developer running three or four coding agents in parallel doesn’t consume 3x the tokens of a chat conversation. It consumes an order of magnitude more. And the subscription price on that seat hasn’t changed.
The exposure most organizations in Iberoamerica aren’t measuring
Over the past two years, teams across the region wove AI subscriptions into their operations. Marketing drafts copy with ChatGPT Plus. Engineering writes and reviews code with Claude Pro. Finance models scenarios. Customer success summarizes tickets. These are no longer experiments — they’re structural workflows.
And most organizations is budgeting AI at current subscription prices.
A 50-person team on Claude Pro costs $1,000 a month. But equivalent API usage for that same team, paying the actual cost of consumed tokens, would be between $15,000 and $40,000 a month depending on usage intensity.
That’s not a rounding error. It’s a budget category that doesn’t exist yet on most P&Ls.
The data confirms it globally. KPMG’s AI Quarterly Pulse Q1 2026 found that U.S. organizations project average AI spending of $207 million over the next 12 months, nearly double the figure for the same period a year prior. But a Goldman Sachs survey found that many large companies are already exceeding their AI budgets by orders of magnitude, with AI spending on track to rival engineer salaries.
Swami Chandrasekaran, director of AI and data labs at KPMG North America, noted that even one or two quarters ago, nobody was tracking LLM consumption costs. Most still aren’t.
The IPO trigger is already in motion
There’s a specific mechanism that’s going to force repricing, and it’s not hypothetical.
Both OpenAI and Anthropic are preparing for IPOs. Anthropic may have exceeded $30 billion in annualized revenue. OpenAI is on track for roughly $25 billion. These numbers look good until you look at the cost side. OpenAI projects $115 billion in cumulative cash burn through 2029.
When you’re private and burning venture capital, you can subsidize inference. An IPO changes the equation overnight. Public markets demand margins. Analysts demand unit economics. Investors demand a path to profitability that doesn’t depend on infinite fundraising.
The fastest way to close the gap between subscription price and actual cost is to raise prices, impose usage caps, or migrate to consumption-based billing. All three options will hit current enterprise subscribers hard.
Signals are already visible. GitHub migrates to usage-based billing on June 1. Microsoft raised Microsoft 365 prices twice in four years, with the latest round specifically tied to AI infrastructure costs. OpenAI launched a $100 Pro plan. Anthropic’s Max plan sits at $200/month. One by one, the floor is rising.
What you should do right now
Three concrete actions — not aspirational, executable this quarter:
1. Audit actual consumption, not seat count. You probably know how many Claude or Copilot seats you’re paying for. You almost certainly don’t know how many tokens those seats are consuming. Get that number. It’s the only way to model your actual exposure when prices adjust.
2. Model repricing scenarios. Apply 2x, 5x, and 10x to your current AI line. If any of those numbers would require an emergency conversation with your CFO, you need to have that conversation now — before you have it under pressure.
3. Build vendor optionality. The organizations that best absorb repricing are the ones that didn’t lock every workflow into a single provider’s stack. Model-agnostic tooling, portable prompts, and documented workflows matter more today than they did 18 months ago.
The era of subsidy isn’t ending because providers want it to. It’s ending because the math always pointed here, and the IPO clock is ticking. Companies that survive the transition will be the ones that treated “AI is cheap” as a temporary condition, not a permanent feature of their cost structure.
The bill is coming. The question is whether your CFO will be surprised when it arrives.