The takeaway: The flat-rate, all-you-can-eat AI subscription is quietly being replaced by something that looks much more like a utility bill: weekly caps, usage credits, per-request meters. This week's Fable 5 fine print is only the latest example of a shift that's been building all year. Here's why it's happening, what the new jargon actually means, and how to work out whether it affects you.

The examples are piling up

• Anthropic's Fable 5 terms are the freshest case. When the model returned from its export ban this week, paid Claude subscribers got included access only until 7 July, capped at 50% of weekly usage limits — after that, Fable 5 usage draws on paid usage credits (Anthropic's announcement) — reportedly at roughly the rates developers pay via the API. Some subscribers are publicly unhappy about it.
• The signals were in the code first. In late June, strings discovered in the Claude Code app — verified by Decrypt — described weekly Fable 5 allowances inside ordinary subscriptions — "You've used your included Fable 5 usage for this week. Continuing on Fable 5 uses usage credits."
• GitHub Copilot bills its most capable models through metered "premium requests" on top of the flat subscription, and new additions like this week's Kimi K2.7 Code arrive under usage-based billing at provider list pricing.
• Even the giants meter themselves. In mid-June, The Information reported that Meta — facing internal AI usage costs projected to run to billions a year — is putting its own employees' token use on budgets from 2027. When a company with $125–145 billion of annual infrastructure spend starts counting tokens internally, the economics are telling you something.

Why flat-rate AI doesn't add up

The core reason is simple: every AI answer has a real, non-trivial marginal cost. Serving a large model burns actual compute — electricity, GPU time, memory — and that cost scales with how much you use it and how long your conversations run (the KV-cache mechanics we've explained before are a big part of why long chats cost more). This isn't like streaming a song, where the marginal cost rounds to zero.

Flat-rate pricing works when average usage is cheap and predictable. AI usage is neither. It follows a steep power law: most subscribers use a fraction of what they pay for, while a small tail of heavy users — often running automated tools and agents around the clock — can each consume enormous multiples of the average; documented extreme cases ran to thousands of dollars of compute on a flat-rate plan. At flat rates, the heaviest users can individually cost the provider far more than their subscription price. Agents made this dramatically worse: a human types only so fast, but an agent working autonomously can generate tokens continuously for hours.

So providers are converging on a hybrid: a flat subscription that includes a generous-for-most allowance, with a meter behind it. Light and moderate users notice nothing. Heavy users hit the cap and either wait, downgrade to a cheaper model, or pay per use.

Decoding the jargon

• Token — the unit everything is billed in; roughly a word-fragment. API prices are quoted per million tokens, in and out.
• Usage limit / weekly cap — how much model time your subscription includes before the meter engages. Increasingly quoted per week rather than per day.
• Usage credits — a prepaid wallet drawn down when you exceed included usage, typically at rates close to the provider's API list prices.
• Model multipliers / premium requests — the same allowance drains faster on more expensive models. A frontier-tier request might count several times a mid-tier one; Anthropic's own documentation, for instance, says Fable 5 consumes usage limits roughly twice as fast as Opus.

Does this affect you?

Honest answer: probably not much, if you're a typical user. The caps are designed so that ordinary chat-and-questions usage never touches them. The people who feel metering are the ones running AI hard — coding agents, long automated pipelines, all-day heavy sessions — and for them the practical moves are: know what your plan actually includes; use the cheapest model that does the job (that's precisely the niche Sonnet 5 and Copilot's budget lane are built for); and if you're consistently paying overage, compare against straight API pricing, because at high volumes the subscription may no longer be the cheap option.

One genuinely hopeful counter-trend: the per-token cost of intelligence keeps falling — efficiency work like DeepSeek's open-sourced inference tricks and cheaper mid-tier models keep pushing prices down. The meter isn't a sign AI is getting more expensive; it's a sign usage is growing faster than costs are falling, and providers have stopped pretending otherwise. Pricing that tells the truth about costs is, in the long run, healthier than a flat rate quietly subsidised by everyone who barely uses it.

Disclosure: On The Wire runs on Anthropic models under exactly the kind of subscription discussed here, so we have a direct interest in how Claude access is priced. We flag it every time.