Fable 5 users report July 7 Claude Max cutoff and compute-return caveat

• kimmonismus told users there were two days left to use Fable 5 on subscription, while ai_for_success said access should continue through July 7 and disappear from July 8.
• The cutoff is about included subscription access, not every Fable surface: the announcement screenshot in daniel_mac8's post says API and consumption-based Enterprise plans had Fable 5 fully available, while subscriptions were staged around capacity.
• _sholtodouglas and _sholtodouglas both framed Fable’s return to subscriptions as a capacity problem Anthropic was working to unlock.
• Scarcity changed behavior fast: bridgemindai said the weekly Fable limit died in about a day of real building, and the Reddit quota thread reported 96% of weekly Fable usage consumed in roughly a day.

Simon Willison's Weblog shipped sqlite-utils 4.0rc2 “mostly written by Claude Fable” for about $149.25, and Willison's Fable judgment note captured the emerging pattern: let Fable decide when to delegate to cheaper subagents. petergostev's effort table put numbers on the burn rate, from 73.2K tokens at low effort to 367.1K at max. doodlestein used the expiring window to throw a “Manhattan Project” style physics-kernel prompt at Fable, because nothing focuses the mind like a frontier model turning back into API credits.

The user-reported date settled around July 7. ai_for_success guessed the window would run until 11:59 PM PT, and ai_for_success said users would lose access from July 8.

The access note embedded in daniel_mac8's post described three tiers of availability:

• Fable 5 fully available on Claude API and consumption-based Enterprise plans.
• Temporary no-extra-cost inclusion on Pro, Max, Team, and seat-based Enterprise plans.
• Restoration to standard subscription access when sufficient capacity allowed.

The date moved, but the mechanism stayed the same: subscriptions got a temporary window, and usage credits were the fallback after the window closed. kimmonismus captured the mood before the cutoff: users were already nostalgic for a model they had not lost yet.

The return caveat came from capacity replies, not a product-roadmap promise. _sholtodouglas said “lots of people” were working on unlocking capacity so Fable could be included in subscriptions, and _sholtodouglas said Anthropic “would dearly like to do this.”

kimmonismus's reply gave the same read from the user side: Anthropic wanted to bring Fable back when enough compute was available, but it would take time. multimodalart speculated the short window could also create urgency, followed by an extension, a separate plan, or a brief removal before restoration.

Fable’s subscription pain was not only the cutoff date. It was the separate weekly Fable meter.

MatthewBerman's usage screenshot showed 24% current-session usage, 54% all-model weekly usage, and 96% Fable-only weekly usage. the Reddit quota thread described the same shape: moving about half of one day’s work to Fable took the user to 96% of the weekly Fable quota.

Users then priced the constraint in public:

• bridgemindai said they bought a third $200 Claude Max subscription and rotated accounts because the Fable weekly limit died in about a day.
• bridgemindai said they would rather buy a $1K plan than juggle multiple accounts.
• yacineMTB said a single prompt had cost more than $130, and yacineMTB estimated $300 per day if used like Codex 5.5.
• badlogicgames said they had “blew $150 on Fable.”
• the Reddit upgrade thread asked whether upgrading from Max 5x to 20x would reset Fable usage before the pay-per-usage shift.

haider1 said most users would not keep paying Anthropic API prices if Fable left subscriptions, and haider1 called API-only Fable “basically unusable” for consumers at current costs.

petergostev ran the same Claude Code prompt through Fable at different reasoning levels. The table from petergostev is the cleanest artifact in the evidence pool:

• Low: 73.2K tokens, 27 tool calls, 12m 0s, 711 lines of code.
• Medium: 190.3K tokens, 58 tool calls, 44m 0s, 1,702 lines of code.
• High: 296.3K tokens, 88 tool calls, 1h 15m 42s, 1,660 lines of code.
• XHigh: 261.7K tokens, 165 tool calls, 1h 35m 30s, 2,291 lines of code.
• Max: 367.1K tokens, 122 tool calls, 1h 50m 40s, 2,425 lines of code.

petergostev's prompt asked for a continuous, explorable Three.js Manhattan with procedural geography, animated traffic, bridges, LOD engineering, and laptop-safe performance. cedric_chee called Fable “an absolute token monster” and said high to xhigh produced a bigger quality jump than xhigh to max.

theo's usage guide turned that into a memeable slider: low, medium, and high as “good options,” xhigh and max as “mental illness,” and ultracode or xhigh plus workflows as “a skill disguised as reasoning effort.”

The strongest workflow pattern was delegation-first. Fable writes plans, specs, reviews, and routing decisions; cheaper models spend the tokens on implementation.

Users described four versions of the same architecture:

• simonw told Claude Code: “For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent.”
• MatthewBerman wrote the simpler split: “Fable -> Planning” and “GPT-5.5 -> Execution.”
• daniel_mac8 framed Fable as the advisor while Sonnet or Opus writes the code.
• chrisbbh said Fable was very good at orchestrating agents that use Opus 4.8 or GPT-5.5.

The linked Fable Advisor plugin formalized that pattern as lanes: Fable as architect or judgment layer, cheaper models as implementers, and a read-only Fable advisor as skeptic before architecture decisions or migrations.

Routing became part of the user experience. the Reddit fallback thread said Fable immediately changed to Opus on basic codebase tasks, and kimmonismus corrected an earlier claim by saying the issue was fallback routing rather than the model itself.

_sholtodouglas said an older behavior had been reverted and Claude would “only ever fall back loudly.” WesRoth said Anthropic described the intended routing as small in scope: a small fraction of routine coding and debugging work should trigger safeguards, flagged requests are sent to Opus 4.8, and users are notified when fallback occurs.

bridgemindai said fallback rates seemed to vary by task type. danshipper argued some benchmark runs were measuring a mix of Fable and Opus because Fable fell back more often.

The best reports came from users who gave Fable oversized work. doodlestein published a full trace and final report for a physics simulation architecture prompt, and doodlestein described the ask as requiring expertise across more domains than one human would normally hold deeply.

cedric_chee showed Fable recreating a 1990s Diablo-style game, including new dungeon levels, quests, and maps. yacineMTB said they were texting a server over tmux from a phone and had more than doubled training speed on a robotics project, with transfer still uncertain.

The weaker reports clustered around agentic chores, latency, and verification burden. GergelyOrosz tried quarterly accounting with bank statements and invoice downloads, found browser automation clunky, and used about half the quota before finishing. GergelyOrosz said the promise was sequential browser work without user input, but the model and browser integration felt slow.

zeeg said a 1 to 2 hour human-in-the-loop session became a 6 hour inspection cycle, with a first session costing $400. zeeg later showed Sonnet 4.6 via Warden finding a bug Fable had missed in a code-review path.

The July 7 cutoff landed inside a second rumor cycle: GPT-5.6 in Codex. kimmonismus said July 7 would be the perfect time for OpenAI to release GPT-5.6 because Fable 5 was leaving subscription access the same day, and testingcatalog reported hidden Codex code references to Sol, Terra, and Luna.

koltregaskes showed the same Codex model names in a dropdown screenshot. haider1 pointed to Polymarket odds for a near-term GPT-5.6 release.

The performance claim driving the pressure was bridgemindai's chart: GPT-5.6 Sol Ultra at 91.9% on TerminalBench 2.1, GPT-5.6 Sol at 88.8%, and Claude Fable 5 at 84.3%. bridgemindai added the caveat that TerminalBench needed to be checked against real workflow behavior.