Metered Into a Ghost Town: Fable 5 Goes Pay-Per-Token and Everyone Defects to GPT-5.6 Sol

Monday the meter starts. Friday’s reset gave every subscriber a full weekly tank of Claude Fable 5 days before Anthropic’s own meter switches on. Sunday is the last free day. From Monday, Fable 5 leaves subscription plans and draws metered usage credits at $10 / $50 per million tokens: roughly double GPT-5.6 Sol, the model that lands one point behind it on the intelligence index.

I’ll make the call now, before it happens: this doesn’t hold. Fable becomes a ghost town, and Anthropic blinks a third time.

The Squeeze Was Already On

Fable’s meter isn’t even the only thing that moves Monday. The same day, the temporary 50% weekly-limit boost Anthropic rolled out in May expires, with nothing announced to replace it. That cushion was load-bearing: it’s been masking a plan that’s tightened for months.

Since the March peak-hour throttle, Max 20x users have watched their weekly quota die by Wednesday, not Sunday. A June bug report documents the limit gone in 2 to 2.5 days even after a June 1 “fix,” the filer cancelling the $200 plan over it. Opus 4.8 made it worse from the other side: the same workflows reportedly burn roughly 3x the tokens they did in spring, so the ceiling didn’t need to drop for the plan to feel smaller. The work got heavier under it.

So Monday is a double cut: Fable priced out, and the inflated ceiling deflated, on a plan that was already emptying faster than it used to. If the squeeze has felt like it tightened week on week, that’s because it did. Monday just takes the padding off, and padding was the last thing holding the goodwill together. Months of mid-week lockouts and quiet reductions have spent the patience people extend to a product they like. Monday asks them to pay more for less, and for a lot of them it’s one squeeze too many.

Why Sol Becomes the Default

A subscription is a different psychological contract from the API. Inside the plan, Fable felt free, so you reached for it by reflex. The moment it draws metered credits, every prompt carries a running cost, and running costs make you hesitate. Hesitation is death for a default.

Metered reads as optional. Optional reads as gone.

And Fable isn’t the only good option anymore. Sol is a rounding error behind it, already inside a subscription the same person probably pays for. So the defection needs no spreadsheet, just a shrug: most people don’t live on the hard-agentic edge where Fable’s 15-point SWE-bench Pro lead shows up, and for the median task Sol is indistinguishable, right there, un-metered. The default flips for the worst reason from Anthropic’s side: not because Sol won on merit, but because Fable started charging and Sol didn’t. And this time the shrug carries resentment. Someone locked out by Wednesday for months doesn’t leave on a cost-benefit calc; they leave because they’re done, and Sol is the open door.

Once it flips, it stays flipped: the muscle memory rebuilds around Sol’s effort dial, its ultra mode, its cache breakpoints, and habits don’t migrate back to a model that now costs extra.

The Plan Runs Out of Reasons

Follow the logic past Fable and it gets worse. Walk the Max ladder as it stands Monday morning:

• Sonnet 5 is sold as the efficient workhorse, and on easy work it is one. On hard work it isn’t: it burns tokens circling the same dead ends, and a problem it loops on can cost more than Opus solving it cleanly the first time. The cheap tier stops being cheap exactly where the price was supposed to matter.
• Opus 4.8 is the flagship you’re actually handing $100 to $200 a month for. On OpenAI’s own price-performance chart, Sol is smarter and cheaper than Opus at every point on the curve. The tier that justifies the subscription is strictly dominated by a rival’s included model.
• Fable 5 is the one model that clears Sol, and it’s the one Anthropic just moved behind the meter.

So the question stops being “will Fable users leave” and becomes “what is the $200 buying.” The single differentiated thing on the plan is now pay-per-token on top of the plan. What’s left inside it is a flagship a competitor beats on both axes. That’s not a churn risk for one model. It’s a churn risk for the subscription.

And for a real slice of serious work, the meter is beside the point: since its redeploy, Fable ships with a tightened safety classifier that refuses or quietly degrades on security, adversarial, and dual-use engineering. Defensible or not, the effect is that the one tier worth paying extra for won’t take some of the hardest problems you’d hand it. OpenAI put this in a launch footnote, needling the model that refuses where the other one ships. Metered and gated: you pay Opus-plus rates for a specialist that can decline the job.

The Lead They’re Risking

Here’s the twist that makes this hubris and not desperation: Anthropic is winning. This isn’t a cornered company lashing out at the meter to survive.

• It just passed OpenAI on business adoption. Ramp’s June index, which tracks what roughly 70,000 US companies actually pay for, put Anthropic at 41% to OpenAI’s 39.5%: a second straight month on top after overtaking in April.
• Its coding foothold was still growing. Claude Code climbed from under 1% of public GitHub commits to around 4% in five months, with credible projections of 20%+ by year end. That’s the curve of a product gaining developers, not shedding them.

You price the flagship out of reach and gate it behind a classifier when you believe the lead is unassailable and the customer is captive. June already showed the belief is wrong: when Anthropic metered programmatic Claude Code usage, power users revolted at effective price hikes some measured between 12x and 175x, and OpenAI answered the same day with two free months of Codex for anyone switching within thirty days. The poach isn’t hypothetical. It’s a standing offer.

The lead is real, which is exactly why I could be early: a moat this wide doesn’t drain in a weekend. But it drains from the edges, and the edges are who the meter hits first.

The Tell Is the Blink History

Here’s why I don’t think Anthropic holds the line: it already hasn’t, twice.

• They extended the free window. Fable’s included-access deadline slipped from July 7 to July 12 in a series of last-hour reprieves. A company confident in the meter doesn’t keep postponing it.
• They reset everyone thirty-one minutes after Sol shipped. That wasn’t a roadmap decision. It was a flinch, and they’ve now trained users to expect the flinch.
• And the position is weaker than when they blinked. Both reversals happened while Opus 4.8 still read as a reason to stay. Monday removes it, and the expensive middle loses the tier above that made its price make sense.

Two blinks and a compounding position is not the setup for a company about to stand firm on its least popular pricing move to date.

The First Receipt

Update, July 13. Within hours of this posting, OpenAI’s product and platform lead put the curve on the timeline: ChatGPT desktop growth, two weeks of ordinary linear climb, then vertical in a single day, squarely inside the metering window. He credits the model and says nothing about Anthropic, which is the gracious way to do it. Good models don’t produce step functions. Pricing events do.

The honest caveat: the y-axis is unlabelled and there’s not an absolute number in sight, so it’s a claim about direction, not magnitude, and it counts arrivals rather than anyone who stayed. But arrivals are the part Anthropic can’t take back. A meter can be reversed in an afternoon. An app already sitting in someone’s dock cannot.

The Whipsaw I’m Betting On

So here’s the falsifiable version. Within a few weeks of the meter starting, one of these happens:

• Fable comes back inside the plans in capped form (a weekly Fable allowance, the way Opus limits worked), or
• The metered price drops to at or below Sol’s, killing the “double the price of second place” problem, or
• The window gets extended again, which is just the blink with a calendar.

Where This Bet Loses

I want the other side on the record, because it’s stronger than the meme version suggests.

• The ghost town might be the target, not the failure. Anthropic has spent two years ending subscription arbitrage; pushing casual users onto Sol while keeping Fable for those who’ll pay is segmentation, not accident. And Fable isn’t dead inside the plan: the orchestrator pattern keeps its judgment in the loop while cheap Sonnet hands do the grunt work. What reads as abandonment may be the pattern they’re forcing.
• Defection isn’t free, and goodwill cuts both ways. Lock-in survives like-for-like migrations: the harness, the memory files, the muscle memory all carry switching cost. And the same users who rage-quit have come back for a reset before, more than once. This market’s loyalty is thin on every side, and a burned user is not always a lost one.
• Opus 4.8 has an edge the chart doesn’t price. Its headline feature was honesty: self-verification a single-shot benchmark score never captures. A team burned once by a confident-wrong answer may keep paying for the model that checks its own work, dominance curve or not.

Each of those is a reason the whipsaw doesn’t come. I still think it does, because none of them survive the one number that matters: a flagship priced at double a rival one point behind isn’t segmentation, it’s a dare.

Metering the flagship, gating it behind a classifier, pricing the plan above the competition: it all rests on one assumption, that the lead is big enough that users have nowhere to go. Sol killed that assumption. The frontier converged to a rounding error, and what’s left is who bills what and who blinks when. Monday, Anthropic starts billing. My money says the blink is already scheduled, because the alternative is watching the exodus make the argument for me. But a reset restores capacity, not trust. You can refill a tank overnight; you can’t refill the goodwill that months of squeezing drained. The blink is coming. It may already be late.

The Third Blink

Update, July 13. Fourteen hours before its own deadline, Anthropic blinked: Fable 5 access on all paid plans extended to July 19, with Claude Code’s weekly rate limits held 50% higher through the same date.

That’s the third bullet above, the blink with a calendar, and it’s the cheapest of the three. Nothing structural moved. Fable’s cap is the same 50% of weekly limits it’s had since the June 30 redeploy, and the credit price on the far side is untouched. Only the date moved, and it moved seven days. The meter isn’t dead, it’s deferred. July 19 is a Sunday, so we do this again next weekend.

The half worth staring at is the other one. That 50% boost to weekly limits isn’t a Fable concession: it applies to every model on the plan, it was sold in May as temporary, and it was scheduled to lapse right alongside the meter. It’s the long-advertised cut, the second blade of the double cut this post opened with, and it was the one thing Anthropic had actually pre-announced and had months to make stick. They couldn’t land that either.

Which is where the capacity story falls apart. Scarcity explains metering Fable: compute is short, so charge for the scarcest thing on the menu. Scarcity cannot explain holding every user’s weekly limits 50% higher across every model, because that spends more compute, not less. The two halves of one announcement point in opposite directions, and only one of them can be about supply. The other is about the door, and OpenAI has already posted a photograph of the door.

Two blinks was a pattern. Three is a policy nobody wants to write down.

The Fourth Blink Is a Settlement

Update, July 18. Two days before its own July 19 deadline, Anthropic called it: from July 20, Fable 5 is “included in all Max and Team Premium plans, at 50% of limits.” No end date, no extension to renew, nothing left to blink about. Pro and Team Standard keep usage-credit access only, softened with a one-time $100 credit.

Scoring the bet honestly: the first bullet hit. Fable came back inside the plans in capped form, an allowance the way Opus limits worked, and the meter died without ever billing a Max subscriber. What the prediction didn’t call is the split. This isn’t a surrender, it’s a settlement: Max keeps the flagship, and the $20 tier loses it for good, because for Pro the “temporary” included window never comes back. The ghost town isn’t averted, it’s zoned. Metered still reads as optional, and optional still reads as gone; it just only reads that way below the $100 line now.

The stated reason is the one you’d expect: “Demand for Fable has been challenging to predict,” capacity secured in stages. Note what the statement doesn’t mention. It landed one day after Moonshot shipped Kimi K3, the first open-weight model with a credible claim on Fable’s tier, while churn receipts were already posting publicly, and with OpenAI’s switching offer still standing. A capacity story that resolves two days early, the morning after the competition arrives, is allowed to be true and still isn’t the whole truth.

And the second blade of the double cut this post opened with is still falling. As of this writing, Anthropic’s support page for the limits promotion is unrevised: “After July 19, 2026, weekly usage limits in Claude Code return to their standard levels.” The announcement that gave Fable back said nothing about the boost, and the arithmetic the tweet didn’t do showed up in the replies within the hour: today’s limits are 150% of base, so reverting is a one-third cut to the ceiling users actually live under. Which makes July 20 a strange kind of gift. Max regains the flagship at 50% of limits, computed against a ceiling a third lower than the day before, on plans where the boosted ceiling was already dying by Wednesday. The settlement made the announcement. The cut didn’t.

Update, July 19. The blade got pulled back the next morning. At 9:04 AM PT on July 18, one day before the deadline, @ClaudeDevs announced the boost now runs “through August 19, for all Pro, Max, Team, and seat-based Enterprise users.” Call it the fifth blink, and note what kind it is: the pre-announced limits cut has now failed to land twice, and the second failure came within a day of the community doing the 33% arithmetic in public. The pattern of this whole saga in one sentence: every deadline Anthropic has set for taking something away has moved when it arrived, and no deadline for giving something back has been set at all. August 19 is a Wednesday. We’ll see.