GLM-5.2 vs Claude Fable 5：差が出たのは推論力だけではなく出力予算だった

この比較は「どちらが絶対に強いか」を決める記事ではありません。実際の API 呼び出しで、GLM-5.2 は出力予算を増やすと数学・物理の推論を正しく返しました。一方で Claude Fable 5 は低い予算でも短く安定し、長い HTML アニメーションではより確実に完走しました。

このテストを見る理由

The test used the Crazyrouter OpenAI-compatible API rather than a chat UI. That matters because the result was not judged only by prose quality. Each response was checked with operational metadata:

Base URL: https://cn.crazyrouter.com/v1
Endpoint: POST /v1/chat/completions
Models: glm-5.2, claude-fable-5
temperature: 0.2
Test date: 2026-07-06

The important fields were max_tokens, completion_tokens, reasoning_tokens, finish_reason, visible content length, whether the generated HTML was closed, and whether the animation actually moved in a browser.

テストした課題

The benchmark deliberately mixed three task types:

Task	Purpose	Reference result
MATH-003	State-based expectation reasoning	Expected flips until HH = 6
PHYS-003	Momentum plus energy accounting	V = 3.0 m/s, x ≈ 0.148 m
CODE-003-ANIM	Long runnable artifact generation	Complete 800x500 Canvas animation HTML

The first two tasks measured reasoning. The third task measured whether a model can produce a complete artifact, not merely a convincing partial code block.

観測結果

Task	glm-5.2	claude-fable-5
Math, original budget	finish_reason=length, completion_tokens=1601, reasoning_tokens=1600, visible body empty	finish_reason=stop, complete and correct
Math, retest	Correct after max_tokens=3200	Retest not needed
Physics, original budget	finish_reason=length, visible body empty	Complete and correct
Physics, retest	Correct after max_tokens=8000	Retest not needed
Animation, original budget	Empty visible HTML at max_tokens=3200	Partial HTML, truncated
Animation, retest	Still truncated at max_tokens=8000	Complete HTML; browser validation passed

The most important observation is that GLM-5.2 was not failing the reasoning itself. In the math and physics tasks, it produced correct answers after a larger output budget. The problem was visibility and completion: a request could return HTTP 200 while the user-facing content was empty or incomplete.

For the long Canvas animation, the difference was sharper. GLM-5.2 produced a visible HTML fragment at max_tokens=8000, but it stopped inside JavaScript and did not close the file. Claude Fable 5 completed the HTML at max_tokens=8000; browser validation showed no console errors, an 800x500 canvas, controls, a speed slider, and changedPixels=55090 after 700 ms.

費用対効果の見方

執筆時点で Crazyrouter の pricing API は glm-5.2 に discount: 0.8 を返しています。つまり、reasoning_tokens と max_tokens をきちんと監視できる用途では、GLM-5.2 はかなり費用対効果の高い選択肢になります。

This is the practical tradeoff:

Workload	Better fit from this test
Short reasoning with enough output budget	GLM-5.2 can be a cost-effective option
Low-budget reasoning responses	Claude Fable 5 was steadier
Long single-file code generation	Claude Fable 5 was stronger in this run
Batch evaluations where metadata is logged	GLM-5.2 becomes easier to operate safely

Do not treat the 0.8 multiplier as a permanent universal price. It is a pricing-data snapshot from Crazyrouter at publication time and should be checked again before a large deployment.

実装時の注意

Minimal request:

curl https://cn.crazyrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $CRAZYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2",
    "messages": [
      {
        "role": "user",
        "content": "Solve the HH expected-flips problem with state equations."
      }
    ],
    "temperature": 0.2,
    "max_tokens": 3200
  }'

To compare Claude Fable 5, keep the same payload and change only the model:

{
  "model": "claude-fable-5"
}

For production-style evaluations, log this shape for every request:

{
  "model": "glm-5.2",
  "max_tokens": 3200,
  "finish_reason": "length",
  "completion_tokens": 3200,
  "reasoning_tokens": 3178,
  "visible_content_chars": 0,
  "html_closed": false,
  "browser_validation": "not_run_incomplete_html"
}

API endpoints should stay clean. Do not add UTM parameters to https://cn.crazyrouter.com/v1. Use tracking only on human-facing article or registration links.

同じ OpenAI 互換リクエストを Crazyrouter で流し、自分のプロンプトで両モデルを比較できます。

FAQ

Did GLM-5.2 fail the reasoning tasks?

No. In this run, GLM-5.2 solved the math task after max_tokens=3200 and the physics task after max_tokens=8000. The issue was that lower budgets were consumed mostly by reasoning tokens before visible content appeared.

Why not score HTTP 200 as success?

Because HTTP 200 only means the API call returned. A benchmark answer can still be unusable if finish_reason=length, visible content is empty, or generated code is incomplete.

Why was the animation task included?

Long code generation exposes a different failure mode. A model can write a convincing first half of a file and still fail if the HTML or JavaScript is cut off.

Is GLM-5.2 still worth testing?

Yes. The current 0.8 discount multiplier makes it attractive for workloads where you can allocate enough output budget and monitor response metadata.

What should be recorded in future comparisons?

At minimum: max_tokens, completion_tokens, reasoning_tokens, finish_reason, visible output length, artifact completeness, and runtime validation.

Final verdict

結論は単純ではありません。GLM-5.2 はコスト面で魅力があり推論も可能ですが、出力予算の管理が必要です。Claude Fable 5 は短い回答と完成した単一 HTML 生成で安定していました。