Posts report Claude Fable 5 included access ends July 7 before $10/$50 token pricing

• trq212's availability post says Anthropic still wants to restore Fable as a standard subscription feature later, but the included subscription window ends after July 7.
• gregisenberg's PSA pegs the post-window credit price at $10 per million input tokens and $50 per million output tokens, matching Anthropic's launch pricing.
• LLMJunky's screenshot of Anthropic's notice shows the caveat that made the relaunch messy: some routine coding/debugging can fall back to Opus 4.8 while the new classifiers are refined.
• LLMJunky's StarSwap run shows the emerging workflow: Fable plans and reviews, cheaper worker models implement, and one Pro session produced 40,000 lines of code.
• AIandDesign's game warning is the counterweight to the hype: one-shot demos can look viral while still missing tuning, difficulty, and finish.

Anthropic's redeployment post says Fable returned July 1 across Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with subscriptions limited to 50% of weekly usage through July 7. The original Fable 5 launch post calls it a safeguarded Mythos-class model and lists the $10/$50 per million-token price. The practical rabbit holes are better: Every's prompt library turns it into a Claude Code delegation machine, while Anthropic's safeguards note explains why some normal-looking coding work can land in Opus 4.8.

Anthropic's redeployment post says Fable 5 is included for Pro, Max, Team, and select Enterprise plans for up to 50% of weekly usage limits through July 7. After that, Fable moves to usage credits.

The Claude usage credits help page says credits are billed separately from subscriptions at standard API rates, with spending caps, alerts, and auto-reload controls. Anthropic's launch page lists Fable 5 at $10 per million input tokens and $50 per million output tokens.

That makes the next few days a weird little holiday for agent nerds. danshipper's poolside post joked that incremental work before the rerelease was basically beach time, while minchoi's unavailable-model screenshot caught the wait right before the switch flipped.

A Hacker News thread focused on the gap between the first promotional window and the restored one: several commenters framed the 50% cap as the bitter part, even if Anthropic says the goal is to bring Fable back to subscriptions when capacity allows.

Anthropic said Fable came back with classifiers designed to target cybersecurity misuse. The same notice said some routine coding and debugging tasks would fall back to Opus 4.8 while Anthropic reduces false positives.

danshipper's fallback note says it is the same model, but benchmarks may measure a mix of Fable and Opus because Fable falls back slightly more. trq212's clarification narrowed the scope to a small fraction of routine coding and debugging tasks.

The wording created the relaunch's first mini-panic. marckohlbrugge's comms thread summarized the loop: people read the announcement as no coding, then an Anthropic team member clarified that coding still works, with fallback risk.

minchoi's Fable-to-Opus meme captured the community read in one image. Funny because expensive models are only funny until the dropdown quietly changes seats.

According to gregisenberg's PSA, Fable usage burns roughly twice as fast as Opus during the included window and only counts up to 50% of a weekly limit. thekitze's token-burn post reported the same pain from subagent-heavy work.

The token sink gets worse in agent-fleet modes. shannholmberg's Ultracode explainer describes Ultracode as a Claude Code mode that fans out agents in parallel, splits the task, adversarially checks findings, and keeps going until the work is covered end to end.

shannholmberg's Fable habits compresses the emerging operating manual into six moves:

1. Give it the why, not just the task.
2. Say what not to do.
3. Let it act, instead of making it over-plan.
4. Force verification.
5. Say less when context and tools are already set up.
6. Do not ask it to explain its reasoning, because that can trip a safety reroute.

petergyang's pre-July-7 workflow gave the shorter version: prep context with cheaper models, plan with Fable, execute with another model, and use lower effort while watching the run. shannholmberg's effort-level reply claimed low to medium effort burned less while still staying above Opus 4.8 x-high in performance.

The cleanest pattern in the evidence is Fable as architect, reviewer, and merge boss. LLMJunky's StarSwap run asked Fable for ten product-improvement proposals, selected seven, had Fable create independent plans, then delegated implementation to GPT 5.5 High through Codex Exec worktrees.

The run broke down like this:

1. Fable inspected the StarSwap codebase and ranked ten improvements.
2. The user selected seven.
3. Fable wrote independent implementation plans into the repo.
4. GPT 5.5 High implemented each plan in its own worktree.
5. Fable reviewed, validated, and fixed the worker outputs.
6. Independent tasks ran in parallel.
7. Fable merged the worktrees and fixed conflicts.

The reported result was 40,000 lines of code using 96% of one five-hour Claude Pro session. The screenshot also shows Fable on High, an approaching session-limit warning, and a 21% weekly Fable limit drawdown.

The pattern lines up with Every's prompt library, which packages 13 prompts for Claude Code workflows around planning, building, verification, overnight delegation, and pull-request prep. The library cites Mike Krieger, Instagram cofounder and head of Anthropic Labs, using Fable for long-running delegated work.

Fable's creative pull is the long-context, tool-using, UI-building loop. MengTo's landing-page demo says Fable understands WebGL, scroll behaviors, and text animations from simple prompts.

viktoroddy's Fable-vs-Opus prompt published the kind of spec people are throwing at it: a React, TypeScript, Vite, and Tailwind v4 landing page with scroll-scrubbed video, GSAP character animation, a pill nav, a glass panel, custom fonts, HLS.js, and mobile behavior.

Other builds were less landing-page-coded:

• bilawalsidhu's IronSight post describes a 4D reconstruction built from two pairs of Meta Ray-Bans, with the old shot-counter roadmap implemented in two nights.
• AIandDesign's dyslexia tutor demo shows a pre-alpha phonics tutor app built for a child with dyslexia.
• rainisto's BeatBandit workflow used Fable through MCP to run screenplay eval agents, inspect BeatBandit's code, find bugs, and tune story-generation prompts.
• techhalla's GTA 2 clone says Fable 5 High created the game in two hours while custom 3D models were made in Tripo.
• GlennHasABeard's platformer update says Fable got a game back up with updated sprites, a local high score, and harder platforming.

The standout creative workflow is not just one-shot generation. It is Fable looking at the artifact, the code, the eval loop, and the product goal in the same run.

The misses are concrete. stevibe's canvas test found Fable smoother than Opus 4.8 on a Mac Studio teardown animation and closer on paper-airplane origami, but neither model folded the plane correctly.

LLMJunky's comparison said GPT 5.5 High produced a better first-try website than Fable, partly because it could generate assets and remove backgrounds. LLMJunky's original Fable website post called Fable's version the ugliest website he had ever seen.

Game builders added a different constraint. AIandDesign's game warning said Fable can produce something that looks cool for a viral post, but it will not automatically tune the game and its difficulty curve into something fun.

AIandDesign's bad-night report later called one Fable session "super disappointing" and ended the night. The expensive model still has dumb hours.