跪拜 Guibai
← All articles
Artificial Intelligence

Claude Fable 5 Leaves GPT-5.5 and Chinese Rivals in the Dust on a Brutal Frontend Test

By 甲维斯 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Frontend code generation from a single complex prompt is a direct proxy for a model's planning, spatial reasoning, and code-synthesis depth. The chasm between Fable 5 and the rest — especially the basic syntax errors and system crashes from models marketed as coding agents — resets expectations about which AI can actually ship a working artifact versus which can only produce a benchmark number.

Summary

Claude Fable 5 produced a richly detailed, code-only cyberpunk rendition of the classical Chinese painting "Along the River During the Qingming Festival," complete with neon signs, flying cars, holographic ads, and interactive shop hover-cards. The generation consumed 63% of a full quota in one go and ran through 38 internal reasoning nodes over 23 minutes. GPT-5.5's web version rushed out an abstract mess in 17 seconds; its Codex variant added more elements but remained crude, with people in the water and a bridge on land. Two leading Chinese models fared worse: one threw a JavaScript syntax error (`=` instead of `:` in an object literal) and displayed nothing, while the other's agent team either refused to send, crashed with an API connection error, or restarted the system. A prior run from that same Chinese model produced a primitive scene where lanterns sat on the ground and figures floated in mid-air, bearing no resemblance to the source painting. The test prompt, all outputs, and conversation logs are published openly.

Takeaways
Claude Fable 5 generated a pure-code, interactive cyberpunk Qingming scroll with neon signs, flying vehicles, holographic ads, and hover-activated shop info cards.
A single generation consumed 63% of the full quota and triggered 38 internal reasoning nodes over 23 minutes, pushing total usage to 112%.
GPT-5.5's web output was an abstract, unrecognizable scene produced in 17 seconds; its Codex variant added elements but placed people in water and a bridge on land.
One Chinese model produced a blank page due to a JavaScript syntax error — using `=` instead of `:` in an object literal.
Another Chinese model's Agent Team refused to send, crashed with an API connection error, or restarted the system; a prior run showed lanterns on the ground and floating figures with no connection to the Qingming scroll.
All tests used the same prompt and were single-run; full conversation logs and outputs are publicly available.
Conclusions

The test is deliberately chosen to be outside the likely training distribution of all models, making it a cleaner signal of generalization rather than memorization.

Fable 5's 38-node reasoning chain and extreme token consumption suggest Anthropic is trading compute for quality at a rate that makes casual use impractical on a free tier.

GPT-5.5's 17-second web response reads as a refusal to engage with the prompt's complexity, not a capability ceiling — the model effectively dodged the task.

A JavaScript syntax error as trivial as `=` for `:` in an object literal indicates the Chinese model's code output is not being validated by even a basic parser before delivery.

An Agent Team that cannot complete a task a single model handles suggests the multi-agent architecture is adding orchestration overhead without robustness, making it strictly worse for this workload.

Concepts & terms
Claude Fable 5
Anthropic's latest flagship model, positioned above Opus 4.8. It exhibits extremely long reasoning chains (38 nodes observed) and high token consumption, producing detailed frontend code from complex prompts.
Cyberpunk Qingming Scroll test
A benchmark prompt requiring a single HTML file that renders a dynamic, cyberpunk-style version of the classical Chinese painting 'Along the River During the Qingming Festival,' with at least 50 animated elements and hover interactions.
From the discussion

The discussion splits between those who see the test as a toy benchmark and those who find the capability gap real. A recurring objection is that headline-grabbing demos don't translate to messy production codebases, where toolchain integration and cost matter more than raw model strength. Others push back, noting that Fable 5's reasoning depth is a genuine step up, though smaller models may still win on cost-performance in agent pipelines.

HTML generation demos are not representative of real-world development, especially legacy Android code or projects using newer APIs.
Launch-day hype consistently overpromises; in complex production projects, these models fall apart.
Model strength alone is insufficient without deep integration into daily dev workflows and toolchains.
Token cost and response speed in multi-turn agent pipelines can outweigh raw capability, making smaller vertical models more practical.
Fable 5 shows a clear lead in reasoning depth, but the gap narrows when paired with strong tooling like ToCodex and DeepSeek V4.
The comparison dimensions in the original article are unclear to some readers, particularly around web vs. desktop output differences.
Featured comments
ToCodex_AI 2 likes

Fable 5 is indeed fierce, but a strong model is one thing — what really matters is whether it can actually land in your daily dev workflow. Yesterday I tried using ToCodex with DeepSeek V4 to write code, and the perceived gap wasn't that big. The key is still engineering pipeline integration. A strong model alone isn't enough; getting the toolchain connected is where the real efficiency comes from.

明略科技

Fable 5 really has pulled ahead in reasoning depth, but benchmark scores and real production performance are still two different things. In many scenarios, a bigger model isn't necessarily better. Token cost and response speed matter more in an agent pipeline, especially for tasks that need multiple consecutive calls. Right now, small models around 4B parameters doing specific vertical tasks actually offer better cost-performance.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗