Anthropic extends Claude Fable 5 access through July 12

• Paid Claude plans got five extra days: Fable 5 was extended through July 12 and included for up to 50% of weekly usage, as shown in LLMJunky's screenshot.
• The cutoff had been July 7 at 11:59:59pm PT, per trq212's clarification, and Fable still appeared as "Included until July 7" in ozansihay's model-picker screenshot.
• Creators used the reprieve for review, planning, and handoffs: shannholmberg's diagram put Fable on the "review and plan" layer, with cheaper models doing execution.
• The sharpest creative workflow treated prompts as production assets: Meng To's Koisei master prompt specified assets, shaders, scroll films, motion language, and performance checks before generation.
• The token math stayed tense: ozansihay quoted post-window billing at $10 per 1M input tokens and $50 per 1M output tokens in his pricing post, while thekitze said he had burned $5K in Cursor credits on Fable.

The in-app copy put the reprieve under a 50% weekly-usage ceiling in Allie K. Miller's screenshot, while the earlier model picker still labeled Fable 5 "For your toughest challenges" in ozansihay's screenshot. The weirdest workflows came from people trying to bank the model's judgment: shannholmberg's recorder wrote learnings back into the repo, Meng To's Koisei prompt treated a landing page spec like source code, and HelloRob's ComfyUI run used Fable as the conductor for cuts, pose/depth refs, image swap, Seedance prompts, and final comparison.

Anthropic extended Fable 5 access on paid Claude plans through July 12. The tooltip in LLMJunky's screenshot says Fable 5 is included in the plan "for up to 50% of your weekly usage limit."

Allie K. Miller saw the same wording and described the change as Fable 5 being extended through July 12 in her post. ClaudeDevs posted a link-only update the same day the screenshots started spreading.

That changed a cutoff users had been planning around. trq212 put removal at 11:59:59pm PT on 7/7 in a clarification, after saying Fable would come off subscriptions after July 7 and return as capacity allowed.

The extension did not make Fable unlimited. minchoi's usage screenshot showed separate weekly meters for "All models" and "Fable," with the Fable meter at 99% used while all-model usage sat at 53%.

trq212 also framed the constraint as capacity, saying "compute can be tricky" while working to make Fable a regular part of subscriptions in a reply to Peter Yang.

Users treated the extension like a compute arbitrage window. thekitze said he had burned $5K in Cursor credits on Fable instead of maxing the $200 Claude subscription before Fable left the plans.

In a follow-up, he said he uploaded passport photos and selfies to buy the 20x plan so Fable could "rule my factory again" until the 12th. levelsio turned the mood into a slogan with "ONE MORE WEEK TO ESCAPE THE UNDERCLASS", and minchoi called it "5 more days."

marckohlbrugge gave the cleanest business framing: expensive tokens need to hit recurring value. His ROI post broke the work into three buckets:

1. Time and token allocation: Fable reviewed git commits, WIP todos, Linear issues, and AI "board meeting" notes, then produced monthly, weekly, and daily planning sheets.
2. Recurring revenue: Fable scoped a job-board Pro subscription and onboarding improvements across apps.
3. Recurring cost cuts: Fable improved an eval suite, tested cheaper models and prompt techniques, found better token caching, and reduced reasoning effort.

He estimated the cost optimization alone would save a few hundred dollars per month in the same post. That is the kind of sentence that makes a $50-per-million output token price feel less abstract.

The most practical post-extension advice was to stop spending Fable on raw execution. shannholmberg's diagram put Fable above the project as a reviewer and planner, then handed execution to Opus or GPT 5.5.

The three runs were concrete:

1. Full feedback loop on what already exists: Fable reads the code and site, then writes the fix plan.
2. Behavior analysis of how the creator works: Fable reviews Claude and Codex sessions, prompts, diffs, kept work, redone work, and loops.
3. Second-brain audit: Fable reads personal and company notes, then maps themes, unused value, gaps, cleanup work, and next builds.

Peter Yang's copy-paste prompt list overlapped with that pattern. His five use cases were: find Fable-worthy work, get life and business advice, make a project ship-ready, plan the next big thing, and refactor an AI skill system.

Yang also linked a tutorial showing the five use cases live.

The strongest playbook treated Fable's output as something weaker models could reuse later. shannholmberg reduced the decision to one question: "can a cheaper model redo this tomorrow?"

The artifacts were specific:

• Workspace standards: Fable rewrites CLAUDE.md, skills, and learnings into conventions, failure rules, and checkable quality criteria, according to shannholmberg's workspace note.
• Consultant audit: Fable audits the business and returns ranked moves, steps, done states, and instructions for a cheaper model in the consultant-audit post.
• Second brain run: Fable researches a niche or competitors, then turns the work into atomized notes, one insight per note, in the second-brain post.
• Goal runs: Claude Code's /goal sets a finish line, while a smaller model checks the stopping condition and the run is capped by turns or time, per the goal-run post.
• Recorder skill: A skill writes a learnings note after each solved problem, leaving reasoning in the repo for later models, according to the recorder post.

The best line came at the end of the thread: whatever gets pulled out keeps working after the model is gone, as shannholmberg put it.

Meng To's one-shot landing page demo

Meng To's Space Coffee demo made the prompt feel like the asset. In his post, he argued that with Fable 5 becoming capable and expensive, the detailed prompt can matter more than building the landing page itself.

The full Koisei prompt in Meng To's follow-up read like a production spec, not a request:

• Stack: Vanilla HTML/CSS/JS or Vite, three.js, GSAP ScrollTrigger, Lenis, and scroll-scrubbed video.
• Assets: exact image and video URLs, preload rules, font choices, color tokens, and texture rules.
• Motion: masked type reveals, WebGL petal particles, water ripple shader, pinned scroll films, horizontal gallery, day-to-night dissolve, and custom cursor states.
• Performance: one shared WebGL renderer, capped device pixel ratio, paused offscreen rendering, lazy video preload, mobile fallbacks, and a 60fps target on an M1 laptop.
• Acceptance checks: smooth loader to hero, reactive petals, noise-threshold dissolve, scrubbed films, adaptive nav, progress rail, and footer wordmark behavior.

Meng To described the workflow as "brief, prompt, html, iterations" in a reply, adding that going to HTML too early costs more tokens. In another reply, he said the work shifts to assets, references, and prompt details, then fine-tuning the prompt instead of repeatedly iterating the final result.

Allie K. Miller's prompting tips pointed in the same direction: review CLAUDE.md against best practices, make skills less prescriptive, and include the why, goals, motivations, success metrics, audience, and ROI.

HelloRob's ComfyUI MCP run

MCP workflows turned Fable into an orchestration layer for creative tools. HelloRob's ComfyUI post said Claude pulled a shot from the web, detected scene cuts, trimmed it, ran DepthAnything v3 and OpenPose workflows, selected the sharpest frame, swapped in a character with gpt-image-2, wrote Seedance 2.0 prompts, generated the clips, and stitched the comparison.

The key line in HelloRob's post was that preprocessing and prompt writing were usually most of the work, and he did not touch a single ComfyUI node.

Magnific used the same urgency to show a Fable 5 plus Magnific MCP thread. The use cases included:

• Developing a 3D game, from Magnific's first example.
• Building a fully interactive responsive website, from the second example.
• Making a product website with 3D objects, from the third example.
• Creating video from a script, from the fourth example.
• Generating four consistent 16:9 luxury skincare campaign mockups, using the prompt in the fifth example.

Character animation got its own pipeline. ai_artworkgen used Midjourney for the original character, OpenAI for character sheet and scene expansion, Dreamina Seedance 2.0 for performance, and Claude Fable 5 for orchestration.

The extension landed after a weekend of demos. minchoi called it probably the most games he had seen vibe-coded from a model.

Min Choi's roundup included a Command & Conquer port to iPhone and iPad in his first example, a Subway Surfers clone in his second, and a kid's toy turned into a mobile game in his third.

The web and 3D side was heavier: an ocean wildlife app in Choi's fourth example, Meta Ray-Bans turned into 4D vision in his fifth, a math-only 3D world in his sixth, a procedural FPS mechanic in his seventh, a live Three.js dungeon generator in his eighth, a dense open world in his ninth, and game-ready Blender assets in his tenth.

levelsio used the window to add drag-and-drop from a local computer into Windows 3.11 on his site. The file saves into C:\DOCS, and shared files go into C:\SHARED, according to his WebFS post.

ozansihay built a browser tool for animated counters that exports WebM or transparent MOV, replacing an After Effects step for one class of edit, as shown in his counter-tool demo.

The practical demos were not all visual. om_patel5's train-platform app post described a UK rail app that predicts hidden platform numbers before stations reveal them, using historical advertised-versus-actual platform data and claiming about 75% accuracy.