Every model launch buries you in the same noise: Twitter threads saying "this just works," demo clips with no numbers, benchmark pages that disagree with each other. When you actually want to know will this work for my task, and what will it cost, the signal is scattered across a dozen places.
So I went through awesome-claude-fable-5, a curated list of 60 real-world Claude Fable 5 cases, each logged with its source, evidence type, and date. The repo's stated principle is the reason I bothered: concrete evidence over hype — reproducible prompts, demos, cost data, and caveats, not vibes.
Here's how it's organized and the cases worth your time.
How each case is logged
Every entry answers four questions, which is what makes the list scannable:
- Input — prompt, screenshot, codebase, logs
- Process — which tool/model, how it was run
- Output — result plus time, tokens, cost
- Evidence type — Demo / Tutorial / Evaluation / Integration
60 cases across 8 categories: coding, agents & long-running automation, games & interactive demos, visual/design/video/3D, documents & research, tutorials & prompt resources, platform/API integration, and evaluations/comparisons/limits.
Coding: screenshot → working clone
One case pastes a single GitHub UI screenshot and asks for a working clone. Reported: ~10 minutes of coding, ~$4.07. The category also covers a single-file macOS-style web OS and large-PR reviews. Good for gauging how far screenshot-to-frontend actually goes today.
Agents: real incident diagnosis, not toy demos
The case that stuck with me: a self-hosted infra outage diagnosed by reading pod logs, querying Cloud SQL error logs, and comparing image digests — landing on "Kubernetes ran a 3-month-old cached image against a freshly migrated DB," then opening a fix PR in minutes, no web search.
Another rebuilt a stalled site (GPT-5.5 and Claude 4.8 had plateaued at 85–90% even with a Figma reference) by going to the original Webflow source, pulling assets, and reproducing it nearly in one pass.
The pattern worth stealing: "Relay"
This recurs across the repo and it's the most practical takeaway: plan, architect, and review with the expensive model; route high-volume implementation to cheaper ones (4.8, GPT-5.5, Sonnet 4.6). One case phrases it as "think with Fable 5, build with a cheaper model, review with Fable 5." If you care about cost, this one idea pays for the read.
The limits are documented too (this is the good part)
The repo doesn't hide the failures. One comparison ran the same Physarum simulation on both stacks: GPT-5.5 (Codex) finished in 17 min for ~$6; Fable 5 (Cursor) took 40+ min and $360.55. A tens-of-times cost gap on the same task. There's also a safety-routing mechanism — cyber/bio-chem/distillation queries get routed to Opus 4.8 instead of refused, triggering in <5% of sessions on average (measured 2–9% depending on the benchmark).
Reproducible prompts you can copy
The tutorials section ships full prompts, not summaries: a DEVICE HARDENING security-pass prompt (model self-check → inventory → 11-category walkthrough → report), a /goal site design-audit prompt with P0–P3 tagging and output paths, and a 4-phase Repo Audit prompt. Worth reading just for prompt-engineering structure.
Trying it yourself
Each case is meant to be reproducible, and the repo includes a curl example against an EvoLink-compatible Messages API:
curl --request POST \
--url https://direct.evolink.ai/v1/messages \
--header 'Authorization: Bearer <token>' \
--header 'Content-Type: application/json' \
--data '{ "model": "claude-fable-5", "max_tokens": 1024, "messages": [{ "role": "user", "content": "Hello, world" }] }'
Reported official benchmarks listed with sources: SWE-Bench Pro 80.3%, Terminal-Bench 2.1 88.0%, OSWorld-Verified 85.0%.
One honest caveat
Almost every number here is creator-reported — pulled from X posts and public demos, curated with its source rather than independently re-verified. The Physarum gap shows how much tool/setup/timing swings results. Treat the figures as a reference range, not gospel.
You can browse the full 60-case list on GitHub (Korean README included).
And if you want to reproduce a case and test the model directly, you can try claude-fable-5 on EvoLink.
Top comments (0)