Usage-based (or “metered”) AI pricing means paying only for what you actually consume — typically the number of tokens an AI model processes — instead of a flat monthly fee that covers unlimited or capped use. Through most of 2024–2025, popular AI products sold flat subscriptions: one price, use as much as you liked within a soft limit. In 2026, several major providers have pulled back from that model for their heaviest workloads, replacing blanket inclusion with metered credits that scale directly with how much compute a task actually burns.
Subscription vs. usage-based, in short
A subscription charges a fixed price per period (say, $20 a month) regardless of how lightly or heavily you use the product. It’s predictable for the customer and simple to market, but it forces the provider to price for an “average” user — anyone far above that average is subsidized by everyone else.
Usage-based pricing flips that: the bill is calculated from metered consumption, most often tokens — the chunks of text an LLM reads and generates, priced per million tokens processed. A short chat question might cost a fraction of a cent; a long, multi-step task can cost dollars. Some products (like Anthropic’s Claude Managed Agents) also meter session runtime, billing per hour a task actively runs.
Most 2026 pricing is actually hybrid: a base subscription still exists, but it now includes a fixed allotment of usage credits rather than unlimited access, with metered charges kicking in once that allotment runs out. This differs from a pure freemium model, which offers a free tier plus optional paid upgrades — here, the paid tier itself has a spending ceiling before extra charges apply.
Why the shift is happening now
The trigger is AI agents — AI systems that don’t just answer one question but run autonomously through many steps, calling tools, writing code, and iterating for minutes or hours per task. Flat pricing was built for a question-and-answer product, where most sessions cost the provider roughly the same to serve. Agentic use breaks that assumption: a single autonomous coding session can consume dozens of sequential model calls and far more compute than a quick chat, yet both were billed identically under a flat plan.
GitHub made this trade-off explicit when it moved GitHub Copilot to token-based “AI Credits” on June 1, 2026, acknowledging that flat pricing had treated “a quick chat question and a multi-hour autonomous coding session” as equivalent costs. Anthropic made a similar move with Claude Fable 5, pulling it out of the usage allowance included in Pro, Max, and Team subscriptions and requiring metered credits instead, at $10 per million input tokens and $50 per million output tokens — Anthropic’s highest published per-token rate. For providers, unmetered heavy use at flat prices is increasingly difficult to sustain against real infrastructure costs; for customers, it means AI bills are starting to track actual usage far more closely than before.
What this looks like in a real bill
Anthropic publishes a worked example: a one-hour coding session on Claude Opus 4.8 using 50,000 input tokens and 15,000 output tokens costs about $0.705 in token charges alone, before any session-runtime fee. Multiply that by dozens of sessions a day and the gap between light and heavy users becomes large — which is exactly the gap flat pricing used to hide. On the Copilot side, developers running extended agentic sessions reported their first usage-based bills landing 10 to 50 times higher than their old flat subscription cost, while casual users who mostly used code completions (which Copilot doesn’t meter) saw little change.
What it means for you
If you rely on an AI subscription for agentic or high-volume tasks — autonomous coding, long research sessions, bulk content generation — check whether your plan still includes unlimited access or has quietly shifted to a metered allotment. Most providers now expose real-time usage dashboards (Anthropic’s Claude Console, GitHub’s billing settings) so you can track spend before a surprise bill arrives; GitHub’s usage-based billing documentation explains how to monitor and cap Copilot credit spend. Lighter, occasional use is often barely affected, since providers still bundle a baseline credit allotment into the subscription price.
In the news
Two recent stories show this shift in practice: Anthropic’s Claude Fable 5 moved from subscription inclusion to metered credits on July 8, 2026, and GitHub Copilot’s first usage-based bills landed the same month, with some agentic-workflow users seeing 10–50x higher costs.
FAQ
Is usage-based pricing always more expensive than a subscription?
Not necessarily. Light users who stay within a plan’s included credit allotment may pay the same or less. It becomes more expensive mainly for heavy, agentic, or long-running workloads that previously benefited from unmetered flat pricing.
Can I predict my AI bill in advance under usage-based pricing?
Roughly, if you know your typical token volume — most providers publish per-token rates and offer usage dashboards. It’s harder to predict for open-ended agentic tasks, since the number of steps a task takes isn’t fixed in advance.
Do all AI products charge this way now?
No. Many consumer AI products still sell flat subscriptions for standard chat use; usage-based metering is spreading fastest where agentic, high-compute features are involved, such as coding agents and autonomous task execution.
How is usage-based pricing different from freemium?
Freemium offers a free tier plus paid upgrades for extra features. Usage-based pricing charges in proportion to consumption — it can apply even on a paid plan, once an included credit allotment is used up.