MoeRouter Blog
English
Get an API key

Claude Fable 5.2 review roundup: the strongest model, the biggest bill, and when it is worth it

Claude Fable 5.2 launched on November 9, 2026 with a 20% price cut. System-card scores, independent tests and Bilibili creators agree: first on general intelligence, and the largest bill. When to use it and when Opus 5.5 is enough.

claude-fable-5-2

In short

  • Claude Fable 5.2 is Anthropic's top model. List price is $8 in / $40 out per 1M tokens: 20% below Fable 5.1 and twice Opus 5.5.
  • The system card reports 93.0% on SWE-bench Pro and 49.1% on FrontierCode 2.0; Artificial Analysis ranks it first on its Intelligence Index with 65.
  • A lower token price did not mean a lower bill: in one hands-on test it wrote about twice the tokens of Fable 5.1 on the same small tasks. Max effort scores below Xhigh and costs three times as much.
  • Reviewers agree it is for the hard problems, with Opus 5.5 as the daily model. On MoeRouter it is offered in the stable and ccmax-cheap groups, not in default.
On this page
  1. What Claude Fable 5.2 is
  2. What it costs on MoeRouter
  3. Benchmarks: first almost everywhere
  4. Hands-on: the token price fell and the bill rose
  5. Weaknesses and things to know
  6. Whether to use it
  7. How to call it
  8. Frequently asked questions

What Claude Fable 5.2 is#

Claude Fable 5.2 is the top model Anthropic released on November 9, 2026. It shares its weights with Mythos 5.2, which is open only to vetted organisations; Fable adds safeguards for cybersecurity and biology[1][2].

Claude Fable 5.2
List price per 1M tokens $8 input, $40 output: 20% below Fable 5.1's $10 / $50
Caching Read $0.20; write $10 (5 minutes) or $16 (1 hour)
Fast mode $16 / $80, up to 2.5x faster
Context and output 1M-token context, up to 128K output tokens
Reasoning effort low, medium, high, xhigh, max. The API defaults to high, Claude Code to medium
For comparison Opus 5.5 is $4 / $20, Sonnet 5.5 $2 / $10
Model name on MoeRouter claude-fable-5-2
Endpoints on MoeRouter Anthropic Messages, Chat Completions

Specifications and prices are as ComputingForGeeks compiled them from Anthropic's material[1]. Anthropic also says output is up to 30% faster than Fable 5.1 and that safeguards step in by mistake about 55% less often[1][2].

What it costs on MoeRouter#

per 1M tokens
ModelGroupInputOutputCache read
claude-fable-5-2ccmax-cheapCN¥9.6≈ US$1.44CN¥48≈ US$7.2CN¥0.24≈ US$0.036
stableCN¥12≈ US$1.8CN¥60≈ US$9CN¥0.3≈ US$0.045

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

Two things to know:

  • It is not in the default group. To call claude-fable-5-2, create an API key in the stable or the ccmax-cheap group; a key of the default group gets model_not_found (the model does not exist or is not offered to that key's group). About groups.
  • The two groups use different routes and cost differently, as the table shows.

Opus 5.5, for comparison:

per 1M tokens
ModelGroupInputOutputCache read
claude-opus-5-5ccmax-cheapCN¥4.8≈ US$0.72CN¥24≈ US$3.6CN¥0.24≈ US$0.036
defaultCN¥5.2≈ US$0.78CN¥26≈ US$3.9CN¥0.26≈ US$0.039
stableCN¥5.2≈ US$0.78CN¥26≈ US$3.9CN¥0.26≈ US$0.039

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

Benchmarks: first almost everywhere#

From Anthropic's system card#

The figures below are as ComputingForGeeks took them from the system card, with Fable 5.2 at maximum effort[1]:

Benchmark Fable 5.2 Fable 5.1 Opus 5.5 GPT-6 Astra
SWE-bench Pro 93.0 81.2 89.9 —
SWE-bench Multimodal 74.1 54.7 61.4 —
FrontierCode 2.0 49.1 32.4 45.5 42.8
FrontierSWE v2 69.9 56.3 62.3 65.5
Terminal-Bench 4.0 67.6 55.8 66.4 57.9
OSWorld 2.1 (offline set) 89.3 83.5 87.2 79.5
Remote Labor Index 41.3 17.9 21.3 20.8
GDPval-AA v2.1 (Elo) 1831 1763 1866 1584

It leads 14 of the 16 results in that table. Where it does not is worth seeing too: Opus 5.5 is higher on GDPval-AA, and the author calls the 1.2-point gap on Terminal-Bench 4.0 noise[1]. Kama Notes adds that SWE-bench Pro is close to saturated, so a three-point gap may not be felt in a real project, and that some of the GPT-6 Astra scores in the table were measured by Anthropic itself through OpenAI's API[2].

From third parties#

  • General intelligence. On the Artificial Analysis Intelligence Index, Fable 5.2 at max effort scores 65 and is first; its max, xhigh and high settings hold the top three places. Gemini 4 Carbon scores 60 and Opus 5.5 at max 59[2]. GPT-6.2 Astra took second place with 62 when it launched on November 17; Fable 5.2 is still first[8].
  • Coding. On Bito's coding benchmark Fable 5.2 scores 83.1%, slightly below Opus 5.5's 85.0% and above GPT-6.2 Astra's 77.5%[7].

More effort is not always better#

Cognition ran FrontierCode 2.0 (132 tasks) in Claude Code at every effort level[1]:

Model and effort Score Average cost per task
Fable 5.2 Low 34.2% $2.10
Fable 5.2 Medium 37.6% $2.48
Fable 5.2 High 45.3% $6.72
Fable 5.2 Xhigh 49.1% $14.19
Fable 5.2 Max 45.7% $44.70
Opus 5.5 Max 45.5% $12.82

Two conclusions. Fable 5.2 at High matches Opus 5.5 at Max for about half the cost. Max scores below Xhigh at three times the cost, so Xhigh is as high as it is worth going. And Claude Code's default, Medium, scores 37.6%: far from the headline number[1].

Hands-on: the token price fell and the bill rose#

ComputingForGeeks compared four models on three one-shot infrastructure tasks: a hardened set of Kubernetes manifests, a defensive Bash log-rotation script and an OpenTofu module for an AWS VPC[1].

Model (medium effort) Output tokens Time Cost
Fable 5.2 10,938 113.5 s $0.446
Fable 5.1 5,510 60.7 s $0.285
Opus 5.5 10,163 86.5 s $0.208
Sonnet 5.5 4,568 38.3 s $0.047

Correctness was effectively a tie between the four. Fable 5.2 wrote about twice the tokens of Fable 5.1, so with a token price 20% lower the three tasks cost 56% more. The author's line: "Claude Fable 5.2 cut the token price and raised our bill." He also found a real improvement: of eight defensive system-administration questions, Fable 5.1 refused two and Fable 5.2 answered all eight[1].

程序员阿江-Relakkes (Bilibili) scored it 94 on his private ModelDial question bank, above Opus 5.5, the previous leader. His description adds that the quota it used "calmed me down", and the video ends on why he still recommends Opus 5.5 for daily use (our translation)[3]. A week later, testing GPT-6.2 Astra, he noted that the same question bank took Fable 5.2 six hours and Astra 94 minutes[9].

kate人不错 (Bilibili) built six new projects in Cursor at Medium effort only, among them a fridge-recognition weekly menu, a Dunhuang mural animation and a brick-building manual, for about $83 at API prices; the video also covers fidelity gaps, interaction bugs and the cost trade-off[4].

盒子姬 (Bilibili) tested it in Godot 4.7 on four tasks chosen to stress low-level ability, rated it top tier and called it the first choice for solo indie games[5].

Dreamtowards (Bilibili) played Saltline, an FPS an Anthropic engineer made with Fable 5.2: the whole game is one 1.4 MB HTML file. According to its developer, the first version started from a one-sentence prompt and took about eight hours and $550[6][9]. The example shows the ceiling and the cost at once.

Weaknesses and things to know#

  • Expensive. Running the whole Artificial Analysis index costs about $26.54 per task on Fable 5.2 and about $4 on GPT-6.2 Astra; on OpenAI's own Agents' Last Exam figures Fable 5.2 costs $60.05 per task[8].
  • Slower than its siblings. It streamed 17% to 25% faster than Fable 5.1, short of the advertised 30%, and both Opus 5.5 and Sonnet 5.5 stream faster[1].
  • Part of its AI-research ability is held back on purpose. Accelerator kernel optimisation and model-architecture work were limited in training to the level of Opus 5.5[2].
  • Max effort truncates easily. At Max it needs a large max_tokens, or it can finish thinking and return nothing[1].

Whether to use it#

Good for

  • Agent runs that last hours, and migrations or refactors across a whole repository
  • Problems other models keep failing (at xhigh)
  • Work where a failure costs more than the tokens

Less suited to

  • Daily coding and code review: Opus 5.5 is half the price and close
  • High-volume, cost-sensitive batch work
  • Short, well-specified scripts

Kama Notes puts it this way: Fable 5.2 is still the model for the hard problems, not the daily one; start with Low and Medium on your real tasks and raise the effort only when they do not get through[2]. ComputingForGeeks says much the same: "Use Fable 5.2 where its lead is real, and let Opus 5.5 carry the rest."[1]

If cost matters, compare with the GPT-6.2 Sol and Astra review: Astra is 3 points lower on the index and much cheaper per task.

How to call it#

Create a key in the stable or the ccmax-cheap group first. In Claude Code:

export ANTHROPIC_BASE_URL="https://moerouter.com"
export ANTHROPIC_AUTH_TOKEN="$MOEROUTER_API_KEY"
unset ANTHROPIC_API_KEY
claude --model claude-fable-5-2

With the Anthropic SDK the base URL is https://moerouter.com, without /v1:

import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://moerouter.com",
    api_key=os.environ["MOEROUTER_API_KEY"],
)
message = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Review this migration plan and point out the steps that can lose data"}],
)
print(message.content[0].text)

It also answers on the Chat Completions endpoint, so a chat client that only speaks the OpenAI format can use it at https://moerouter.com/v1. More in calling with an SDK. The docs are in Chinese for now.

Frequently asked questions#

Claude Fable 5.2 or Opus 5.5?#

Opus 5.5 for daily coding and code review: its list price is half of Fable 5.2's and the gap is small on most tasks. Fable 5.2 for agent runs that last hours, repository-wide migrations and problems other models cannot get through.

How much does the Claude Fable 5.2 API cost?#

List price per 1M tokens is $8 input and $40 output, with cache reads at $0.20: 20% below Fable 5.1. MoeRouter's reference prices are in the table above; they differ by group, and accounts are charged in Chinese yuan.

Why does my key say claude-fable-5-2 does not exist?#

Because the model is not in the default group. Create a key in the stable or the ccmax-cheap group in the console, or change the group of the key you have, and call it again.

How high should I set Fable 5.2's reasoning effort?#

Start at Medium or High. On FrontierCode 2.0, High matches Opus 5.5 at Max for about half the cost. Go no higher than Xhigh: Max scores lower and costs three times as much.

Is Fable 5.2 better than GPT-6.2 Astra?#

Fable 5.2 scores 65 and Astra 62 on the Artificial Analysis Intelligence Index, and Fable 5.2 is higher on Bito's coding benchmark too. Astra does the same work far more cheaply and quickly: on a Bilibili creator's private question bank the two were one point apart.

They are the same model weights. Fable 5.2 adds safeguards for cybersecurity and biology and is open to everyone; Mythos 5.2 has no such safeguards and is open only to vetted organisations.

How do I use Fable 5.2 in Claude Code?#

Set ANTHROPIC_BASE_URL to https://moerouter.com, use a key of the stable or ccmax-cheap group, and start with --model claude-fable-5-2, or set ANTHROPIC_MODEL to it in settings.json.

Sources

The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.

  1. Claude Fable 5.2: Benchmarks, Pricing and Real Tests ComputingForGeeks · published 2026-11-11
  2. Claude Fable 5.2发布:降价20%、1000个Agent形式化费马大定理,和Opus 5.5、GPT-6 Astra、Gemini 4怎么选 (in Chinese) 卡码笔记 (Kama Notes) · published 2026-11-12
  3. Claude Fable 5.2实测:94分刷爆了我们的私有题库,但我真用不起了! (video, in Chinese) Bilibili · 程序员阿江-Relakkes · published 2026-11-10
  4. 最强模型6项实测!Fable 5.2 只用 Medium,就能做到这个程度 (video, in Chinese) Bilibili · kate人不错 · published 2026-11-10
  5. 【盒子姬】Godot 中 Fable5.2 能力实测 (video, in Chinese) Bilibili · 盒子姬 · published 2026-11-12
  6. [品鉴] Fable 5.2 一句话生成仅 1.4MB 的 FPS 游戏,画质却如此惊艳?!AI 只用了 8 小时 (video, in Chinese) Bilibili · Dreamtowards · published 2026-11-11
  7. GPT-6.2 Sol and Astra benchmark: how they compare to Claude Opus 5.5, Fable 5.2 and GPT-6.1 Sol for coding Bito
  8. GPT-6.2 Astra and Sol: Features, Benchmarks, Tests & Pricing DataCamp · published 2026-11-17
  9. GPT6.2 Astra全网首测:93分硬刚Fable 5.2,屎山题库直接给我秒了? (video, in Chinese) Bilibili · 程序员阿江-Relakkes · published 2026-11-12

Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.