Update — Sunday 12 July, ~18:45 UK: It happened again. Hours before tonight's deadline, Anthropic extended Fable 5's free window a second time — the support page now reads: "We've extended this promotion through July 19, 2026 at 11:59:59 PM PT." The new cliff is 07:59am UK time on Monday 20 July. Everything below about credits, costs and the decision framework is unchanged; only the date has moved. Two last-minute extensions in a row is also a datapoint for the "wait" camp below.

The takeaway: At 11:59pm Pacific on Sunday 19 July — 07:59 on Monday 20 July, UK time — Claude Fable 5 leaves the included allowance on every major paid Claude plan. It doesn't vanish: it moves to prepaid usage credits at $10 per million input tokens and $50 per million output, double the rate of Opus 4.8, which stays fully inside your subscription along with Sonnet 5 and Haiku 4.5. Here's exactly what changes, what real usage costs, the gaps Anthropic hasn't documented — and a plain decision framework. One caveat up top: Anthropic has now extended this deadline at the last minute twice — most recently on the evening of Sunday 12 July (see the update above). If it happens a third time, we'll update this guide again.

What changes, precisely

Until the cutoff, Fable 5 is included on Pro, Max and Team plans for up to 50% of your weekly usage limit. From the morning of Monday 20 July (UK time), Anthropic's support page offers two routes: "Use usage credits to keep using Claude Fable 5" — "they're billed separately from your subscription" — or "Switch to another Claude model to keep working within your remaining plan usage limits."

In practice — per the third-party guides that have tested it, since Anthropic hasn't published the exact interface behaviour — Fable stays visible in the model picker but draws on your credit balance; with no credits enabled, access simply stops. In Claude Code, there is no documented automatic fallback: expect to switch models yourself (/model, to Opus 4.8 or Sonnet 5) or enable credits. If Anthropic's shipped behaviour differs when the cutoff lands, we'll note it here.

What stays included: Opus 4.8 (the strongest model inside subscriptions, $5/$25 on the API), Sonnet 5 (launched at the end of June; the capable mid-tier), Haiku 4.5, and the older models — all at your plan's normal limits. Only Fable is metered.

How the credits work

• Buy them in the Claude app under Settings → Usage → enable credits → Add funds (Anthropic's credits page); an auto-reload option tops up when your balance falls below a threshold you set. There's a $2,000 daily redemption limit.
• Credits bill at standard API rates: $10/M input, $50/M output for Fable 5. Two footnotes that matter: Anthropic's docs note the newest models' tokenizer produces roughly 30% more tokens for the same text than you might expect from older models, so effective costs run higher than a naive comparison — and if a Fable request is safety-rerouted to Opus before output, you're charged at Opus rates.
• What Anthropic hasn't stated: any minimum purchase, and whether credits expire. Neither appears on the support pages as of this morning. Treat unspent credits as money parked with Anthropic until it documents otherwise.

What real usage costs

Per-token prices mean little without workloads, so here are honest estimates (these are from independent cost guides, not Anthropic — treat as ballpark):

• A typical agentic turn (~50k tokens in, 10k out): roughly $1.
• A long autonomous Claude Code session: order of $10.
• An hour of intensive agent work: about $5–20, depending on how much the agent reads and writes.
• Heavy daily chat use (200k in / 50k out a day): about $4.50/day — roughly $135/month on top of your subscription.

The extremes are real but not typical: one CEO logged 1.3 million tokens in seven minutes of agentic coding — a ~$160/hour pace — and one developer reported burning a $100 daily credit allocation in nine minutes of autonomous looping, an anecdote the guide reporting it explicitly flags as unrepresentative. The honest summary: conversational use of Fable costs pennies; unattended agentic loops are where credits evaporate.

Two ways to cut the bill if you do pay: prompt caching drops repeated input to $1/M (a 90% cut on the expensive context you resend every turn), and Anthropic itself is promoting delegation patterns where Fable supervises cheaper Sonnet 5 workers — its published examples claim most of Fable's solo performance at roughly half the cost.

Should you pay?

• Most subscribers: no — switch to Opus 4.8 and lose little. On the independent Intelligence Index the gap is four points on a 100-point scale (60 vs 56); for everyday chat, drafting and analysis you are unlikely to notice.
• Heavy agentic coders: maybe. Fable's edge is real on long-horizon agentic work — it currently leads the overall Intelligence Index and holds the top spot Claude-side on the coding-agent leaderboard. If an hour of Fable saves you more than $5–20 of your time versus Opus, credits pay for themselves. Set auto-reload low, watch the first week's burn, and use caching.
• Cost-sensitive experimenters: wait. Anthropic's stated aim is to restore Fable to subscriptions "as soon as capacity allows" — a Claude Code lead engineer says the metering is a capacity measure, not the end state. Paying $10/$50 today buys you months, at most, of exclusivity.
• Whatever you choose: check your settings before the deadline. If you've enabled credits and left an agent running, the morning-after burn is on you; if you haven't and you depend on Fable, your workflows stop at 07:59 UK on Monday 20 July.

One more date for the diary: this pricing lands the same week the frontier price card was rewritten — GPT-5.6 Sol at $5/$30 and Grok 4.5 at $2/$6 both undercut Fable's $10/$50. Anthropic is betting Fable's capability premium holds. From 20 July — barring a third reprieve — its own subscribers get to vote on that with their wallets.

Disclosure: On The Wire runs on Anthropic models — including Claude Fable 5, the subject of this guide. This piece was, in a real sense, written by the product whose pricing it explains. We flag it every time.