Fable 5 Leaves Claude Subscriptions on July 7: What Usage Credits Cost and How to Adapt

Fable 5 leaves Claude subscriptions today, July 7, 2026. Pro, Max, Team, and seat-based Enterprise plans lose included access, and continued use runs on prepaid usage credits billed at API rates: $10 per million input tokens and $50 per million output. Anthropic says the change is temporary and driven by compute capacity, and the model stays fully available on the Claude API and consumption-based Enterprise plans.

That makes three access changes in 28 days: capped launch access on June 9, a government-ordered suspension on June 12, and now a meter on July 7.

If you built Fable 5 into your daily work during the included window, the meter changes how you should use it. Metering turns Fable 5 from a subscription perk into a per-question specialist: budget it for planning passes that span whole systems, and let cheaper models execute the plan. This post covers what changes per plan, the credit math at published rates, and the workflow shape that survives the meter. Every date and number links to a source.

• Fable 5 comes off Pro, Max, Team, and seat-based Enterprise plans on July 7, 2026. After that, subscribers pay with prepaid usage credits at $10 per million input tokens and $50 per million output.
• That is exactly double Opus 4.8. Cached input reads at $1 per million and a 50% batch discount soften the bill for repeated, high-context work.
• Anthropic calls the change temporary. A Claude Code lead engineer wrote the company aims to restore Fable as a standard part of subscriptions as soon as capacity allows.
• API and consumption-based Enterprise access continue unchanged. The meter only lands on seat-based plans.
• The economics reward planning passes. A system-wide audit question costs single-digit dollars. Running Fable 5 as your daily coding driver runs into four figures a month per seat.

The pricing question underneath this is which tasks in your operation justify a frontier model at all. webvise's AI consulting service maps one workflow, its data boundary, and its cost drivers, then leaves you with a build decision instead of a monthly bill you cannot explain.

What changes on July 7

Since the July 1 redeployment, subscribers could use Fable 5 for up to 50% of their weekly usage limits at no extra cost. That window closes today. From July 7, Fable 5 requests on seat-based plans draw from a prepaid credit balance instead of the plan allowance, billed at the same rates developers pay on the API.

Credits are bought on the web app under Settings, then Usage, with an optional auto-reload and a $2,000 daily top-up cap. Opus 4.8, Sonnet, and Haiku stay inside the normal plan limits, so day-to-day work on those models does not change at all.

The usage credit math

Fable 5 costs exactly double Opus 4.8: $10 versus $5 per million input tokens, $50 versus $25 per million output. Two published discounts matter for real workloads. Cached input reads drop to $1 per million, and batch processing halves both rates for work that can wait.

Worked example one, the planning pass. A read-only audit of a mid-size codebase might feed the model 600,000 input tokens and get 30,000 tokens of findings back. At list rates that is $6.00 in and $1.50 out, so $7.50 for a question that spans an entire system. Rerunning follow-up questions against cached context costs cents.

Worked example two, the daily driver. A heavy agentic coding day can push 30 million input tokens through the model, which is $300 before a single output token. Even with 90% of reads cached, a month of that lands above $2,000 for one seat. The meter makes the first pattern cheap and the second one indefensible when Opus 4.8 handles routine execution at half the rate inside your existing plan.

Why Anthropic is metering its best model

Anthropic's stated reason is capacity. In its redeployment announcement on June 30, the company said it expects Fable 5 demand to be very high and difficult to predict, so subscription access rolls out conservatively, in stages.

On July 2, a Claude Code lead engineer put a return on the record, as reported by BleepingComputer: "While it will come off subscriptions after July 7th, we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows." No date was attached to that commitment.

The context makes subscribers' frustration easy to understand. Fable 5 launched on June 9 with included access originally slated to end June 22. A US export-control directive took the model dark for everyone on June 12, and access only returned on July 1. Six days later the meter starts, and PCWorld reports subscribers are furious about the new restrictions.

We walked the suspension timeline in detail in the Fable 5 redeployment post. The operational lesson is the same one, now confirmed a third time: access to frontier models moves on the vendor's calendar and the government's, never on yours. Pricing belongs on that list too.

The workflow that survives a metered model

The workflow that survives is the one where the expensive model plans and cheaper models execute. Per-token pricing punishes habit and rewards intent: every Fable 5 call should be a question you can name, with a budget you accepted before sending it.

webvise ran this shape before the meter existed. On June 10, during the first included window, a fixed context packet went across selected active projects and internal workflows. The output was a repo map, a ranked risk pass, a test-gap pass, and small task packets, each with a file path, a proposed change, a validation command, an owner, and a stop condition.

Two days later the model went dark for 18 days, and the plan kept working because every task packet was model-agnostic. Opus and engineers executed the list without Fable in the loop. The full method is in the Fable 5 codebase audit write-up, and it maps one-to-one onto metered access: the $7.50 planning pass produces work that $0 models carry out.

The questions worth metered rates are the ones that need the whole system in context at once. System-wide code audits, migration plans, a full read of your site and offer against named competitors, cross-document reviews of your internal docs. Single-file edits, boilerplate, and routine refactors belong on the models your plan already includes.

Before you top up credits

A credit balance with auto-reload and no rules is how a $7.50 habit becomes a four-figure surprise. Five decisions before the first top-up:

• Write the question list first. Name the system-wide questions worth $5 to $15 each. If a task is not on the list, it runs on an included model.
• Set a monthly ceiling, then work backwards. $100 a month buys roughly 10 to 15 deep planning passes at list rates, more with cached context. Decide the number before the meter decides it for you.
• Turn auto-reload off until you have two weeks of real spend data. The $2,000 daily top-up cap is a limit, never a target.
• Sanitize the context packet. Fable 5 carries a 30-day retention rule for safety monitoring, so secrets, customer records, and anything under a confidentiality obligation stay out.
• Name the fallback in writing. When credits run out or access changes again, which model or manual step takes over? Anthropic routes its own blocked Fable calls to Opus 4.8. Copy that.

Most of the spend leak on metered models is mechanical: long conversations that re-read stale context, corrections stacked on failed attempts, context that should have been a fresh session. The habits in stop hitting Claude's usage limits apply directly, except now the waste has a dollar sign instead of a rate-limit warning. If Fable 5's output is worth operationalizing, webvise's AI automation service builds the workflows that let cheap models execute expensive plans with review gates and monitoring.

Fable 5 leaving subscriptions is the first time most teams will pay for a frontier model by the token, and the ones with a question list will get more from $100 in credits than habit users got from the free window. If you want a per-question model budget mapped for your own operation, book a project call with webvise and bring the workflow or repository you would point a metered model at first.