Anthropic extends paid Fable 5 access through July 12 with 50% weekly cap
Anthropic says paid Claude plans keep Fable 5 access until July 12, with Fable use capped at 50% of weekly limits. Users still reported exhausted quotas and multiple Max subscriptions as workarounds.
TL;DR
- Paid Claude plans keep Fable 5 through July 12, and Claude's announcement says the 50% weekly Fable cap still applies before usage credits kick in.
- The extension added days without refilling spent buckets; aibuilderclub_ put it as “you get the days, not the tokens” in one post, while Simon Willison's screenshot showed Fable already at 100% for the week.
- Anthropic's developer account quickly reframed Fable usage around delegation: ClaudeDevs said Sonnet 5 plus a Fable advisor hit about 92% of Fable's SWE-bench Pro score at about 63% of the price.
- The demand was not theoretical: Simon Willison's writeup says sqlite-utils 4.0rc2 was mostly written by Fable for about $149.25, while doodlestein's FrankenSim update showed a multi-crate Rust numerics stack moving under Fable-heavy Claude Code sessions.
- The access fight is now a pricing and routing story, with AlphaSignalAI's roundup listing Fable at $10 input and $50 output per 1M tokens and noting that some flagged cyber and biology queries route to Opus 4.8.
Anthropic's support article now sits beside the earlier rollout post that said Fable would leave subscriptions after July 7 unless capacity allowed an extension. The strangest artifact is the UI copy: Willison's popup told users Fable was included through July 12, while the same usage panel showed his Fable bucket empty. Christmas came early for coding-agent nerds, then the receipt arrived.
July 12 access
Anthropic said Claude Fable 5 will remain available on all paid plans through July 12. The follow-up from the same account says users can spend up to 50% of their weekly usage limit on Fable 5, then continue with usage credits or switch to another Claude model.
Rohan Paul captured the support-page wording: Pro, Max, Team, and premium seat-based Enterprise users get promotional access where enabled, with the promotion extended through July 12, 2026 at 11:59:59 PM PT in his screenshot.
The extension reversed the immediate cutoff users had been planning around. trq212 had said Fable would come off subscription plans after July 7 and return as a standard subscription model “as soon as capacity allows” in the earlier timeline, then clarified the exact prior end time as 11:59:59pm PT on July 7 in a reply.
50% weekly cap
The policy kept Fable's separate weekly bucket. Willison found the extension from a Claude web UI popup, then posted a usage panel showing 53% all-model weekly usage and 100% Fable weekly usage, both resetting Wednesday at noon in his follow-up.
Reset timing created uneven outcomes. testingcatalog said their reset “happened at the right time,” while most people appeared to have different weekly reset days in one account-level observation.
The user mood flipped from relief to accounting because many had spent the quota before the extension. dexhorthy first said everyone they knew had burned the allowance before the weekend, then corrected that the 50% weekly bucket does reset in the same thread. haider1 framed the complaint more sharply: Anthropic redeployed Fable on July 1, held the 50% cap through July 7, then extended without a reset after users had maxed out in his post.
Quota economics
Fable access turned subscriptions into capacity arbitrage. bridgemindai said four $200 Claude Max subscriptions, $800 per month total, bought roughly one full week of Fable because weekly limits died in a day in his usage post.
Theo estimated he got about $2,267 of Fable usage over six days across two $200 accounts in his calculation. Steve Yegge said six of eight “mule accounts” had already exhausted Fable tokens until July 11 and joked that ten Claude Max accounts were back on the menu in a reply.
Claude Code's usage-credit flow made the overage visible inside the tool. giffmana posted a terminal-style menu showing $73.50 spent against a $70 monthly limit, with a “buy more” option directly in the workflow in his screenshot.
Zeeg pushed back on the token-spend flex: a $1,000 per week per person habit comes out to roughly $50,000 per person per year, which he compared to adding 15% to 25% more engineering headcount in the cost thread. That was the cleanest economic read in the pile.
Advisor and orchestrator patterns
Anthropic's developer account published the workaround as an agent pattern: keep Fable in the judgment path and let cheaper models emit most of the tokens.
The official patterns were concrete:
- Advisor: Sonnet 5 executes, calls Fable 5 for guidance, and bills most tokens at the executor rate per ClaudeDevs.
- SWE-bench Pro result: Sonnet 5 plus a Fable advisor gets about 92% of Fable's score at about 63% of the price, with Fable called roughly once per task according to ClaudeDevs.
- Orchestrator: Fable 5 plans and delegates to Sonnet 5 workers, so token-heavy execution runs on the cheaper model in the second strategy.
- BrowseComp result: A Fable orchestrator with Sonnet 5 worker sub-agents gets 96% of Fable performance at 46% of the price in ClaudeDevs' BrowseComp post.
- Sub-agent cache: Claude Managed Agents gives each sub-agent its own cache, so repeat calls do not pay the full context cost twice according to ClaudeDevs.
Community tooling moved in the same direction. daniel_mac8 described fable-advisor as a Claude Code plugin that lets Fable choose Opus 4.8, GPT-5.5, or Sonnet 5 as the implementer, then perform a read-only review pass at the end in the plugin post.
Fable work that made the cap hurt
sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)
I wrote about the sqlite-utils 4.0rc1 release a couple of weeks ago. Since we only have Claude Fable on our Max subscriptions for a few more days, I decided to see if it could help me get to a 4.0 stable release that I felt truly comfortable about, since I try to keep to SemVer and like my incompatible major versions to be as rare as possible. I started with this prompt, in Claude Code for web on my iPhone: Final review before shipping a stable 4.0 release - very important to spot any last minute things that would be a breaking change if we fix them later Here's that initial report it created for me. There were some significant problems that I hadn't myself encountered yet - 5 that Fable categorized as "release blockers". Here's the worst of the bunch: 1. delete_where() never commits and poisons the connection (data loss) Table.delete_where() (sqlite_utils/db.py:2948) runs its DELETE via a bare self.db.execute() with no atomic() wrapper — compare Table.delete() at db.py:2944, which wraps correctly. The connection is left in_transaction=True, so every subsequent atomic() call takes the savepoint branch (db.py:430-440) and never commits either. Reproduced end-to-end: db = sqlite_utils.Database("dw.db") db["t"].insert_all([{"id": i} for i in range(3)], pk="id") db["t"].delete_where("id = ?", [0]) # conn.in_transaction is now True db["t"].insert({"id": 50}) db["u"].insert({"a": 1}) db.close() # Reopen: rows are [0, 1, 2] — the delete, row 50, AND table u are all gone. That's a re
Willison used Fable to push sqlite-utils 4.0rc2 toward a stable major release. His writeup says Fable found five “release blockers,” including a delete_where() transaction bug that could leave the connection in a state where later atomic operations never committed, losing the delete, later inserts, and a new table on close in the RSS item.
doodlestein went broader. He asked Fable to explain the sheaf cohomology parts of FrankenSim back to him, then used a scoring rubric to produce a plan addendum in his process post. Later screenshots showed Fable-heavy sessions landing complex Rust numerics work, including a deterministic complex eigensolver, cross-ISA golden hashes, and an Orr-Sommerfeld acceptance path in the FrankenSim update.
Fable was also being spent on demos and games. Cedric Chee said he had Fable recreating a 1990s Diablo-style game with generated dungeon levels, quests, and maps in his build post, while skirano posted a Fable-made NES-style FIFA World Cup 2026 demo in his video post.
GPT-5.6 timing
The extension landed while users were already treating GPT-5.6 as the next pressure point. Rohan Paul said GPT-5.6 Sol, Terra, and Luna would become public Thursday, with preview access expanding worldwide in his post.
That timing shaped the competitive read. haider1 argued OpenAI's best release window was exactly when Fable was being pulled from subscriptions, because demand would move if most users lost access in his post. bridgemindai went further, saying Anthropic could not afford to remove Fable from subscriptions if GPT-5.6 landed in Codex at the same time in his post.
A TerminalBench 2.1 chart circulating in the same thread placed GPT-5.6 Sol Ultra at 91.9%, GPT-5.6 Sol at 88.8%, Claude Mythos 5 at 88.0%, and Claude Fable 5 at 84.3%
.
Routing rules
AlphaSignalAI's model roundup supplied the missing production detail: Fable returned through Claude Platform, claude.ai, Claude Code, and Claude Cowork after a June 12 suspension and July 1 restoration, with a listed price of $10 input and $50 output per 1M tokens in the Fable item.
The same roundup says Fable uses the same underlying model as Mythos 5 with cyber and biology safeguards, and that many flagged cyber and biology queries route to Opus 4.8 instead. Users are not charged Fable prices for rerouted requests, according to AlphaSignalAI's summary.
That routing layer explains why subscription availability, safety handling, and perceived model quality collapsed into one argument. QuixiAI said Fable had slowed down and was “playing dumb” on kernel work in one complaint, while giffmana said none of his Fable chats had been downgraded in a contrasting post.