Usage Credits

What It Is

A usage credit is a unit of pre-paid or subscription-bundled spend that a model platform applies against your inference bill. The customer-facing number is dollars; the underlying mechanic is token-priced API calls drawn down from a personal credit-pool. Anthropic’s June 23 move on Claude Fable 5 is the cleanest example of the pattern. Pro and Max subscribers no longer get unlimited Fable 5 access within their plan. They get a credit allowance to spend on Fable 5 specifically, and once the allowance is gone they top up at the API rate of ten dollars per million input tokens and fifty dollars per million output.

The shift matters because it dissolves the old promise of an AI subscription. For two years the consumer pricing pattern was “twenty dollars a month, unlimited within fair use.” That bundle worked when the cost per query was low and a $20 plan could swallow a heavy user’s monthly token spend. Frontier models broke the math. A single one-million-token Fable 5 call now costs Anthropic more than a Pro subscription pays in a month. Credit allowances are how labs ration the new top tier without breaking the rest of the subscription.

How It Actually Works

You buy a plan with a credit budget attached: Copilot Pro now bundles $15 of credit, Copilot Max bundles $200, Claude’s bundled Fable 5 allowance is structured the same way. Every call against the metered model decrements the credit balance at a published per-token rate. When the balance hits zero, the platform either falls back to a cheaper model automatically, blocks further calls until the next billing cycle, or lets you top up at the standard API price. The choice belongs to the vendor and varies by plan.

The mechanic has two consequences operators feel immediately. First, your spend becomes lumpy and unpredictable in a way it was not when the plan was flat-rate. A heavy reasoning week can burn through the credit pool by the tenth of the month and leave you unable to use the top model for the next twenty days. Second, the per-call cost is now legible to the user in a way it never was before. When the credit balance is showing on screen, every prompt has a price tag, and the operator starts asking whether the prompt is worth it. That second consequence is the more important behavioral shift.

Why It Matters Right Now

Three vendor moves in three weeks made usage credits the default consumer billing primitive of mid-2026. GitHub Copilot transitioned all plans to credit-based billing on June 1. Anthropic shipped Claude Fable 5 with a thirteen-day free window that ended June 23, after which the model moves to credit-metered access at the API’s token-pricing rate. Google’s $250 Ultra Gemini tier is the holdout, but the Ultra subscription buys Deep Think reasoning at fixed cost only at the cap; beyond a usage threshold, the same metering applies. Operators who built workflows on the assumption that the best model was bundled into a flat plan need to rebuild the cost forecast.

The Cost / Tradeoff

The honest tradeoff: credit pricing makes top-tier inference accessible at smaller scale than enterprise contracts ever did, but it makes the cost asymmetry visible. A workflow that mixes ten Fable 5 calls with five hundred Opus calls used to cost the same as a workflow that mixed two Fable 5 calls with five hundred Opus calls. Today the first one costs five times as much. The cleaner an operator’s intuition about which task actually needs the top model, the better credits work for them. The fuzzier the intuition, the faster the budget evaporates.

How TWO Uses It

The Wise Operator runs the model-routing call directly: the cheapest model that can hold the work is the right model. We treat usage credits as a budget for genuinely hard tasks: a research question with no clear surface answer, a contract clause whose interpretation matters, a piece of code whose edge cases will hurt if missed. Every other prompt routes to a cheaper tier under a simple model-routing discipline. The mental shortcut is to ask, before pressing send, whether you would spend three dollars to know the answer you are about to ask for. If the answer is no, you should not be asking the model on the credit-metered tier; switch to the bundled tier and rewrite the prompt to fit it. The discipline keeps the credit pool around for the prompts where the answer actually matters, and it bends the usage-based-pricing curve in your favor rather than the vendor’s.

A Concrete Operator Scenario

A two-person consulting team subscribes to Claude Max with a $200 monthly credit allowance and uses Fable 5 for client-facing strategy memos. In the first week of the new pricing, both partners run their normal workload through Fable 5 and burn $140 of credit on memos that an Opus 4.8 run would have handled. By day twelve, the pool is empty and the team is back on Opus 4.8 for the rest of the month with no Fable 5 access for client-final work. The fix is not to buy more credit. The fix is to use Opus 4.8 for the first draft and Fable 5 only on the second pass when the memo’s framing actually needs the harder model. That single workflow change recovers about two-thirds of the credit budget.

What to Watch Next

The signal that tells you usage credits are maturing as a primitive is when vendors publish per-task price calibrations rather than per-token rates. Today’s pricing forces the operator to convert from prompt length to dollars in their head. The next iteration will price by job type: a contract review costs you $X of credit, a research brief costs you $Y. When that happens, the credit budget becomes a planning tool rather than a usage anxiety. Watch for it from whichever lab needs the next pricing differentiator, and watch the small print on the auto-fallback rules: the vendor that quietly drops you to the cheaper model when your credits run low has effectively turned credits into a soft cap rather than a hard one.