Metered or Included? The AI Pricing Divide Is a Bet on Who Absorbs the Variance

On June 30, Anthropic shipped Claude Sonnet 5 as a cheaper way to run agents, priced at an intro $2 per million input tokens. Days later, it moved Claude Fable 5, the model developers had been leaning on for the heavy work, off flat subscription limits and onto metered usage credits billed at full API rates. The same week, OpenAI reaffirmed that Codex stays included on every ChatGPT tier, with overage packs optional. Two market leaders. Opposite billing philosophies. Announced almost simultaneously.

For a while, the story everyone told (us included) was that flat-rate agentic coding was dead and everyone was moving to the meter. When Copilot, Cursor, and Anthropic all switched inside a year, we argued the flat-rate era was ending and read it as an industry converging on one answer. That read is now incomplete. The picture this summer is not convergence. It is a fork.

What is the difference between metered and included AI pricing?

Included pricing bundles agent usage into a flat subscription, and when you hit the ceiling you wait for the reset. Metered pricing charges usage credits at token rates once you cross an allowance, so a heavy agent loop bills straight through. The real difference is not the sticker. It is who absorbs the variance of an unbounded workload.

Look at the three biggest players and you see three different answers. OpenAI holds Codex included on every ChatGPT tier. Anthropic meters Fable 5 after an included window (as of publication the cutover lands July 7, though Anthropic has walked pricing seams back inside a heat window before). Google bundles model access into its plans while quietly debiting a compute budget behind the sticker. Included, metered, and bundled-but-debiting are not three prices. They are three positions on one question.

And the question is not "how much should a token cost." It is "who eats the difference when an agent loop runs ten times longer than expected." Metered puts that on the builder. Included puts it on the vendor. The quiet-debit bundle hides it until the allowance runs dry. Same underlying volatility, three different owners.

Why does AI coding cost so much more some months?

Because agentic coding spends tokens per action, and both the number of actions and the model tier vary enormously run to run. The bill is not just high. It is unpredictable. Ramp's June 2026 AI Index, built on real corporate-card payment data, found business AI spend swinging a median of about 58% month over month. The variance is the story, not the average.

That number sits on top of a longer trend. Ramp's payment data shows business AI spend up 497% year over year on usage up 1,001%: people are not paying a little more, they are doing vastly more work through the models. Treat those figures as directional rather than universal (Ramp reads a self-selected book of mostly US startups and mid-market firms), but the shape is unmistakable. This is the usage drift we called structural, not a pricing gimmick, showing up now in settled invoices.

Here is the paradox that makes it counterintuitive. Over roughly the same period, the price per token fell about 280x. Cheaper tokens, bigger and swingier bills. The cost driver is behavior (how many tokens an agent burns per action), not unit price, which is exactly why a flat subscription is so hard to hold: you are promising a fixed price for a workload with no fixed size.

The variance also has a direction. Ramp found the share of AI spend going to premium model tiers flipping from about 5.7% to 55.9% in under a year, even though premium models are only around 46% of tokens. The expensive work is migrating up-tier.

That flip is why the divide will widen, not close. As the premium share climbs, "included" gets more expensive to honor and "metered" gets more punishing to receive. The two bets pull further apart the more capable the models get.

Why would a rational vendor pick each side?

Because each side is a defensible bet, not a moral stance, and the honest read respects both. Included is a scale bet: OpenAI can pool variance across an enormous user base and treat Codex as retention. Metered is a margin bet: Anthropic protects unit economics and lets the heaviest users pay their weight. Neither is naive. Neither is greed.

The quiet-debit bundle is a third bet, on perception: keep the sticker flat and throttle silently, so the customer feels a plan rather than a meter. All three are rational responses to the same volatility. They just disagree about who should be surprised by it, and how visibly.

This is the same move we watched Anthropic make when it split its own bundle, where the seam a vendor draws is the contestable part, not the fact that they drew one. What is new is that the biggest players are now drawing the seam in genuinely different places, and that difference is a signal about how each one reads its own risk. It is worth noting this lands in a broader lock-in week: the pricing divide is the economic half, and the technical half is your tools paying a schema tax at the same time. The frontier's convenience has both a bill and a seam, and builders own both.

Should I choose a metered or included AI coding plan?

It depends on who can absorb the variance, because this is a build decision, not a spectator sport. Picking included caps your downside but throttles you at the ceiling. Picking metered removes the ceiling but passes through every spike. The choice of vendor is now partly a choice of who your finance team argues with.

But the builder-side answer is the same on either side of the divide: manage the variance yourself. Simon Willison's discipline this month is the clearest version of it. Reserve the expensive model for judgment, and delegate the routine implementation to cheaper tiers. Intentional aim beats prompt-and-pray precisely because it controls token behavior, which is the one variable the billing model cannot fix for you.

That is what economics-of-building looks like in the agentic era. The pricing seam is a product constraint you design around, the same way you design around latency or a rate limit. Teams that treat token behavior as an engineering variable, not a surprise on the invoice, win regardless of which billing model their vendor picks.

So the divide is not a phase on the way to one settled model. It is a standing fork, and which side you can live on tells you something real about how well you understand your own agent loops. If a 58% monthly swing would break you, you do not have a pricing problem. You have a variance problem, and no billing model fixes that for you.