Three days from now, Fable 5 stops being something you already paid for.
If you’re on Claude Pro, Max, Team, or a seat-based Enterprise plan, Fable 5 has been included in your subscription since July 1 — capped at 50% of your weekly usage limit, but included. On July 12, 2026 that ends. The model stays in your picker. It just starts billing against prepaid usage credits at API rates: $10 per million input tokens, $50 per million output tokens.
Most of the coverage you’ll find on this got the date wrong, because Anthropic originally set the cutoff at July 7 and pushed it back five days after subscribers made enough noise. The date isn’t the interesting part anyway. The interesting part is that almost nobody has run the arithmetic on what a Fable 5 session actually costs, which is the only number that decides whether you should turn credits on.
What changed, in order
June 12: US export controls landed and Anthropic suspended Fable 5 and Mythos 5 for everyone. I wrote about that suspension and what it revealed about model-weight policy at the time.
June 30: the controls lifted. July 1: Fable 5 came back globally, with a new constraint — paid subscribers could burn it for up to half their weekly usage cap, no extra charge, through July 7.
July 7: Anthropic extended that window to July 12 after the announcement went over badly. Same terms, five more days.
After July 12, included access ends for Pro, Max, Team, and seat-based Enterprise. Premium Enterprise seats had a slightly better deal during the window (Fable drew from each member’s seat usage) and land in the same place. Anthropic has said it wants Fable 5 back as a standard subscription model once capacity allows, and has attached no timeline to that. Treat “no timeline” as “not this quarter.”
The billing mechanics people are getting wrong
You will not wake up on July 13 with a surprise charge. Usage credits are opt-in and prepaid, and that distinction matters more than the price.
You enable them under Settings → Usage, click Add funds, and put money in. If your subscription came through the Apple or Google app store, you can only do this on the web — the mobile apps won’t let you buy credits. Auto-reload is available and off by default: set a reload amount and a threshold, and Anthropic tops you up when the balance drops below it. Set $20 and $5, and your card gets charged $20 whenever the balance dips under $5. There’s a $2,000 daily redemption ceiling on top of everything.
So the failure mode isn’t a runaway bill. It’s the opposite: you don’t enable credits, you pick Fable 5 out of habit on July 13, and you get told you can’t use it. Which, honestly, is the correct default for most people.
What a session actually costs
Here’s where every other post on this topic stops and mine keeps going. $10/$50 per million tokens means nothing until you know how many tokens a session eats.
A few things to hold in your head. Output tokens cost 5x what input tokens cost, so verbose models and long generations dominate the bill. And a chat resends the entire conversation on every turn, so a long thread’s input cost grows roughly quadratically with turn count, not linearly.
A long chat. Forty turns, conversation growing to around 60k tokens by the end, averaging maybe 30k tokens of context resent per turn. That’s about 1.2M input tokens. Say 600 output tokens per reply — 24k output.
- Input: 1.2M × $10/M = $12.00
- Output: 0.024M × $50/M = $1.20
- Total: about $13.20
An agentic coding session. Two hours, lots of file reads, a few dozen tool calls. Call it 2M input tokens and 150k output.
- Input: 2M × $10/M = $20.00
- Output: 0.15M × $50/M = $7.50
- Total: about $27.50
A document analysis run. A 300-page PDF is roughly 150k tokens. Ask five questions against it, 1k output each.
- Input: 0.15M × $10/M = $1.50
- Output: 0.005M × $50/M = $0.25
- Total: under $2.00
Now the caveat that changes those first two numbers, and that I have not seen a single other write-up mention. Fable 5 supports prompt caching at a 90% discount — cache reads bill at $1/M instead of $10/M. On the API, a long chat’s resent prefix is exactly the thing caching is built for. If claude.ai applies caching to your credit-billed sessions, that 40-turn chat’s input cost falls from $12.00 to somewhere around $2.30, and the coding session lands near $13 instead of $27.50.
Whether it does apply, Anthropic hasn’t published. So the honest range for a heavy Fable 5 chat is $3 to $13, and for a real coding session $13 to $28. Watch your balance drop over your first session and you’ll know which world you’re in within twenty minutes. Anyone quoting you a single confident number for this hasn’t checked.
Batch API pricing is $5/$25, half rate, if you’re doing anything asynchronous. And US-only inference runs at 1.1x.
The number that reframes all of this
Claude Pro is $20 a month. Max runs to $200.
Two agentic coding sessions a week on Fable 5, at the un-cached rate, is roughly $220 a month in credits — on top of your subscription. Even at the cached rate it’s over $100. You are not topping up a subscription at that point. You are paying API prices with extra steps, and you should think about it the way you’d think about an API budget.
That’s the whole decision, really. Fable 5 through credits isn’t a subscription perk anymore. It’s metered inference wearing a subscription’s clothes.
Fable 5 versus the alternatives, at July 2026 prices
| Model | Input /M | Output /M | Coding session (2M in, 150k out) |
|---|---|---|---|
| Fable 5 | $10.00 | $50.00 | $27.50 |
| Opus 4.8 | $5.00 | $25.00 | $13.75 |
| Opus 4.8 Fast Mode | $10.00 | $50.00 | $27.50 |
| Sonnet 5 (intro) | $2.00 | $10.00 | $5.50 |
| GPT-5.6 Sol | $5.00 | $30.00 | $14.50 |
| GPT-5.6 Terra | $2.50 | $15.00 | $7.25 |
Opus 4.8 is exactly half of Fable 5 on both sides of the meter, and it’s included in your plan. That’s the comparison that should be doing the work in your head. Fable 5 has to be twice as useful to you as Opus 4.8, per token, to break even — and then it has to clear that bar against a model you’re already paying for, so the real hurdle is higher than 2x.
Sonnet 5’s introductory pricing runs through August 31, after which it goes to $3/$15. Sonnet 5 also ships Anthropic’s newer tokenizer, which produces roughly 30% more tokens for the same text, with the exact hit depending on what you’re feeding it. That reads like a hidden tax on the cheap model until you check which models share that tokenizer: Fable 5 and Opus 4.8 are both on it too. Everything in the top half of the table counts your text the same way, so the ratios between them hold. What it does mean is that every absolute number in this post runs about a third higher than a Sonnet 4.6-era estimate would have told you. I compared Sonnet 5 against the previous generation in more depth here.
Opus 4.8’s Fast Mode is priced identically to Fable 5, at $10/$50. Worth knowing before you assume the premium tier is the expensive option. It isn’t; it’s the same money for faster output on a model that costs half as much in normal mode.
So what’s Fable 5 actually for
Nothing in that table says Fable 5 is bad. It says Fable 5 is expensive, which is a different claim, and the price only tells you what to do once you know what you’re doing with it.
Long-horizon work is where I’ve found it earns the multiple. Tasks where a single dropped constraint twenty steps back poisons the output, and you’d rather pay 2x than re-run the whole thing and re-verify it by hand. Deep research passes across a dozen sources. Refactors that span a lot of files with interdependent invariants. The common thread is that the cost of being subtly wrong is much higher than the token bill.
Where it doesn’t earn it: anything you’d catch immediately if it went wrong. Drafting, summarizing, routine code, answering questions you already know the shape of. For that work, Fable 5 is Opus 4.8 at double the price, and half of what you’re buying is confidence you didn’t need.
The framing I keep returning to — and which applies well beyond this one model — is that the frontier tier is insurance against verification cost, not a general upgrade. When verification is cheap, buy the cheaper model. Same logic I’ve argued in the model routing guide, and it hasn’t stopped being true.
What to do before Friday
If you use Fable 5 occasionally, for hard problems. Enable credits, put in $20, leave auto-reload off. Twenty dollars buys you a fair amount of episodic use and it can’t quietly become two hundred. Check the balance after your first real session — that’s your actual per-session rate, and it beats any estimate in this post, including mine.
If Fable 5 is your daily driver. Move to Opus 4.8 for a week before the deadline forces you to. Not after. You want to find out whether the quality gap is real for your work while you still have the included access to A/B against, rather than discovering it on July 13 with a credit balance draining in the corner of your screen. My honest guess is that most daily-driver users are on Fable out of habit, not need. Prove me wrong on your own workload.
If you’re a team lead with seats to think about. Credits are per-user, so “just enable it for everyone” is an unbounded line item with no per-seat cap. Enable them for the two or three people whose work genuinely justifies frontier-tier inference, hold the rest on Opus 4.8, and revisit in a month with real numbers instead of vibes. Seat-based Enterprise plans are in the same boat as Team here, whatever your rep implied during the June outage.
If you’re building on the API. Nothing changed for you. API pricing was always $10/$50 and still is. This whole story is about the subscription apps.
The part that should make you a little uneasy
Anthropic says Fable 5 returns to plans when capacity allows. Maybe it does.
But look at the shape of the move. A frontier model shipped inside a subscription, then pulled out and metered, with the fallback being a cheaper model that’s genuinely good. It works because Opus 4.8 is good enough that most people won’t miss Fable. That’s not a temporary capacity story, that’s a pricing architecture — and it’s the same one GitHub Copilot landed on with AI credits, and the one I sketched out in the agent pricing piece back in June. Flat subscription for the commodity tier, metering for the frontier. Everyone’s converging on it because inference costs at the frontier don’t amortize the way software costs do.
Which means the useful skill isn’t picking the right model this week. It’s knowing your own token volume well enough to price any of these offers in under a minute.
Go enable credits, put in $20, and run one session you’d normally run. Whatever number comes back is the only one in this post that’s about you.
Sources: Anthropic — Redeploying Claude Fable 5, Claude Help Center — Manage usage credits, Claude Platform Docs — Pricing, Android Authority — extension after backlash, OpenAI API Pricing