The Fable Tax

Tomorrow night — July 12, 11:59 PM PT — Claude Fable 5 stops being included in the Claude subscription. On July 13 it flips to metered usage credits: $10 per million input tokens, $50 per million output tokens, billed on top of whatever you already pay for Pro, Max, or Team. No credits enabled, no Fable. Enterprise seats don’t even get a grace allowance.

Anthropic already blinked once. The cutoff was set for around July 7. The backlash was loud enough that they moved it — to tomorrow. The official line is that they “hope to eventually restore Fable 5 as a standard benefit once capacity allows.” No timeline. That’s the situation. Here’s my read as a Max subscriber, and what I’m actually doing about it.

The clock runs out tomorrow

Let’s be fair to the model first. Fable 5 is Anthropic’s Mythos-level model — the most capable thing they sell — launched June 9 and built to run for days at a time inside an agent harness like Claude Code. It is genuinely excellent. I’m not disputing the model. I’m disputing the bill.

The timeline: free promotional inclusion on Pro, Max, Team, and select Enterprise plans (up to 50% of your weekly limit) was set to end around July 7. After backlash, Anthropic pushed it to July 12 at 11:59 PM PT. Metered billing starts July 13. After that Fable is not included — you enable usage credits and pay API-style rates, and Enterprise seats won’t run it at all without credits turned on. Anthropic says it hopes to restore Fable to subscriptions “once capacity allows,” with no date attached.

The numbers that matter

Anthropic lists Fable 5 at $10 per million input tokens and $50 per million output tokens (with a 90% prompt-cache discount on input, a 1.1x multiplier for US-only inference, and a 30-day data-retention requirement). At those rates it is one of the most expensive widely available models on the market. Here is the ladder — Fable 5 is Anthropic’s official pricing; the other Claude rungs are reported figures, and the OpenAI family is from launch coverage:

Two things jump out. Fable is exactly twice Opus 4.8 on both input and output. And Opus stays included in the subscription while Fable walks out of it.

Then there’s the other column of that table. OpenAI shipped the GPT-5.6 family on July 9 — Sol, Terra, and Luna, plus a “ChatGPT Work” agent. Sol lists at $5 / $30: half Fable’s input rate, 60% of its output rate. Sam Altman says Sol is 54% more token-efficient on coding, and it runs up to 750 tokens/sec on Cerebras. Put plainly: Anthropic is putting its flagship behind a second paywall the same week a cheaper frontier rival became generally available. Sol went GA July 9. Fable’s meter starts July 13. Four days.

Why the subscription math breaks

Metering by itself isn’t the problem — frontier inference costs money, and paying per token is a coherent deal when I *choose* the API. My objection is narrower: I won’t pay a metered flagship premium on top of a flat Max subscription.

I bought Max because it’s a flat deal — one price, use the models. The moment the best model comes out of that deal and onto a per-token meter, I’m paying a subscription *and* an API bill for the one model I upgraded to reach. That’s not a subscription anymore. It’s a subscription with a turnstile in the middle of it.

And $50 per million output tokens is brutal on a model designed to run autonomously for hours. Long agent loops are exactly where output tokens pile up — so the meter punishes the thing Fable is best at. The prompt-cache discount doesn’t repair that logic, and “it might come back someday” doesn’t either. The flagship is leaving the plan tomorrow with no return date. So, plainly: I’m not paying it. Not because Fable isn’t worth it — because paying twice for one tier breaks the deal.

My stack after midnight

Here’s the actual configuration as of tomorrow:

• Opus 4.8 stays the daily driver. It’s included, it’s half Fable’s price, and it’s still a frontier model. For ~95% of what I do, Opus is not the bottleneck.
• Sol becomes my main agent. It’s GA now, cheaper at both ends of the token stream, and claims a real efficiency edge. Cheaper per token *and* fewer tokens is the whole ballgame for agentic work.
• Opus is the second opinion. It keeps Claude in the loop as a reviewer without accepting Fable’s second tollbooth.

And if Anthropic keeps its flagship gated behind a meter? Then I look hard at whether I need Max at all, versus Pro plus Sol. I’d rather not — the flat-subscription value is the reason I’m loyal. Metering the flagship is Anthropic putting that value on the table.

Anthropic blinks again

Here’s where I’ll stick my neck out: Anthropic folds Fable 5 back into the subscription — or heavily softens the metering — sooner than “when capacity allows” suggests. That’s my opinion; the evidence under it is fact.

First, they already blinked. The July 7 → July 12 extension wasn’t generosity, it was a response to backlash — which tells you the pressure works and they know it. Second, and bigger: Sol. A week ago there was no credible, cheaper frontier model to walk to. Now there’s one, it’s four days old, and it undercuts Fable on both price and efficiency. The capacity constraint is probably real — Fable is expensive to serve — but “we can’t afford to give it away” collides hard with “our competitor just gave away something almost as good for half the price.” That collision usually goes to the side with the cheaper alternative.

Hot take: Fable’s metering isn’t a premium strategy. It’s an acquisition campaign for Sol — paid for by Anthropic.

Competition has the leverage now

None of this is a complaint. It’s the opposite. A year ago, if Anthropic metered its best model, you shrugged and paid — where else were you going? Today you have somewhere to go, and that single fact rewrites the negotiation. Fable might be the better model. Sol being *good enough at half the price* is what keeps Fable honest.

Tomorrow the meter starts. For the first time, the customer holds a card that reads “or I’ll just use the other one.” The deadline belongs to Anthropic. The verdict belongs to the subscriber. Mine is no.