Belin Doc IconBelin Doc

Claude Fable 5.2 for Translation: Where It Adds Value, Tested on 10 Document Scenarios

BelinDoc Team2026/11/11

What Claude Fable 5.2 brings to translation: whole-document context, the strongest multilingual understanding in the Claude family, fewer mistaken refusals, and clean results across 10 real document scenarios.

Key takeaways

  • Fable 5.2 handled all 10 of our document scenarios cleanly: consistent terminology, untouched code and tables, and OCR errors corrected without being asked.
  • Its 1M-token context and 128K-token output let it take in a whole long document at once, which is what keeps terms consistent from the first page to the last.
  • It has the strongest multilingual understanding in the Claude family: 96.0% on Global MMLU across 42 languages, ahead of Claude Opus 5.5 at 94.3%.
  • Anthropic reports about 55% fewer mistaken safeguard blocks than Fable 5.1, which matters for medical, legal and security documents.
  • It costs $8 / $40 per million tokens, twice Claude Opus 5.5, so it pays off on long, complex or high-stakes documents rather than routine ones.

Quick answer: Claude Fable 5.2 is worth it for translation when the document is long, technical or high-stakes. It can take in a whole long document at once, it has the strongest multilingual understanding of any Claude model, and in our 10-scenario test it kept terminology consistent, left code and tables untouched, and corrected OCR errors on its own. For short, routine texts, start with Claude Opus 5.5 at half the price, which is also Anthropic's own advice.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Most coverage focuses on coding and long-running agents. We translate documents, so this review asks one question: what does Fable 5.2 change for translation?

Where Fable 5.2 adds value for translation

1. A whole long document in one pass. Fable 5.2 reads up to 1M tokens and writes up to 128K tokens in a single response. Translating a long contract, manual or thesis as one piece, instead of in disconnected chunks, is what keeps a term translated the same way on page 3 and page 300.

2. It gets ambiguous terms right from context. In our test, "bus" and "driver" appear across three paragraphs, and only the last paragraph says that "bus" means the internal I2C bus, not a message bus. Fable 5.2 used 总线 (bus) and 驱动程序 (driver) consistently throughout, never confusing a hardware driver with a person driving.

3. It repairs messy scans instead of copying the errors. Our OCR passage contained "rnodule", "be-fore", "devi ation", "O.5 %" and "inspecti0n". Fable 5.2 translated the intended words (module, before, deviation, 0.5 %, inspection) without being told the text had OCR errors.

4. Strongest multilingual understanding in the Claude family. On Global MMLU, which covers 42 languages from French and German to Yoruba, Igbo and Chichewa, Fable 5.2 averaged 96.0%, ahead of Claude Opus 5.5 (94.3%), Fable 5.1 (94.0%) and Claude Sonnet 5.5 (92.1%). That's a measure of understanding rather than translation, but understanding the source is where translation starts, especially for less common languages.

5. Fewer mistaken refusals on sensitive material. Anthropic reports about 55% fewer safeguard blocks than Fable 5.1 in early testing (about 60% fewer for cybersecurity and 40% for biology). For translators of medical, pharmaceutical and security documents, that means fewer jobs interrupted by a false alarm.

6. Cheaper and faster than the last Fable. It costs 20% less than Fable 5.1 and generates output up to 30% faster, which adds up on long documents.

What's new in Claude Fable 5.2

All figures are from Anthropic's announcement and the Fable 5.2 / Mythos 5.2 system card.

Claude Fable 5.2
ReleasedNovember 9, 2026
Price$8 per million input tokens, $40 per million output tokens (Fable 5.1 was $10 / $50)
SpeedOutputs up to 30% faster than Fable 5.1; optional fast mode up to 2.5× faster at $16 / $80
Context window / max output1M tokens / 128K tokens
Knowledge cutoffJune 2026
ThinkingAlways on (adaptive)
AvailabilityGenerally available, including on Amazon Web Services, Google Cloud and Microsoft Azure

Claude Mythos 5.2 is the same underlying model with lighter safeguards. It's limited to organizations in Anthropic's Life Sciences and Cyber Verification Programs, at the same price.

Anthropic's own recommendation: start with Claude Opus 5.5 ($4 / $20), which it calls its strongest daily driver, and move up to Fable 5.2 for the most ambitious work.

Anthropic's headline demonstration is that a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, in a single run per model.

What the benchmarks say about translation

Short version: not much, and two numbers are easy to misread.

BenchmarkFable 5.2Claude Opus 5.5What it actually measures
Global MMLU (GMMLU)96.0%94.3%Multiple-choice knowledge questions in 42 languages
MILU94.1%93.1%Knowledge questions in 10 Indic languages plus English
SWE-bench Multilingual96.6%93.9%Software engineering tasks in nine programming languages
FrontierCode 2.049.1%45.5%Agentic coding
  • Global MMLU is the most relevant result. Fable 5.2 is the strongest Claude model at understanding questions written in many languages. But the model only picks an answer; it never writes a translation, so fluency, terminology and formatting aren't measured.
  • SWE-bench Multilingual isn't about human languages. "Multilingual" here means programming languages, so it can't support any claim about translating documents.
  • No translation benchmark was published. The system card says Claude "is multilingual and typically responds in the same language as the user's input. Output quality varies by language."

So any claim that Fable 5.2 improves terminology consistency or localization is an inference, not a published result. That's why we ran our own test.

Our test: 10 real document scenarios

How we ran it

  • Material: the eight scenarios from our GPT-5.6 translation review, plus two new ones: a Chinese contract clause into English, and a medical passage into Chinese.
  • Instruction: the same translation instruction we used for GPT-5.6, asking the model to keep code, identifiers, numbers and paragraph breaks and to output only the translation.
  • Environment: one run of Claude Fable 5.2 through Claude Code, on November 11, 2026. Not the raw API, so effort and sampling settings weren't adjustable.
  • Review: we checked every output by hand against the source. There was no judge scoring in this round, so these results aren't comparable with the scores in our earlier reviews.

What we found

ScenarioWhat we checkedResult
Academic abstract (EN→ZH)Terms, O(n log n), percentagesAccurate; notation and numbers intact
Software license clause (EN→ZH)One long legal sentence, the exception at the endAccurate; the "except to the extent such restriction is expressly prohibited by applicable law" carve-out kept
Technical doc with code (EN→ZH)cache.ttl, cache.strict_mode, file path left untranslatedAll code kept exactly
Lu Xun, My Old Home (ZH→EN)Rhythm and imageryNatural literary English ("beneath a sallow sky")
Manga dialogue (JA→EN)Shōnen character voiceFitting voice ("This ain't over yet!")
Markdown table (EN→ZH)Table structure, E_RETRY, 250 ms, HTTP 507Structure, codes and units intact
Long-document consistency (EN→ZH)"bus" and "driver" across three paragraphsConsistent throughout: 总线 (bus) and 驱动程序 (driver), never mistaken for a message bus or a person driving
OCR noise (EN→ZH)rnodule, be-fore, devi ation, O.5, inspecti0nAll recovered correctly
Contract clause (ZH→EN)Contract register"shall", "first installment", "liability for breach of contract"
Medical (EN→ZH)Clinical termsStandard terms for insulin resistance, β-cell dysfunction and first-line therapy

We found no omissions, additions or broken formatting in any of the ten outputs.

The catch: short passages can't separate top models

These passages are short, from about 80 to 430 characters. Fable 5.2 handled all of them cleanly, and so did the strongest models in our July review, which averaged about 4.7 out of 5. On material like this, frontier models bunch together near the top, so a test of this size can't show whether Fable 5.2 translates better than Claude Opus 5.5.

Where a stronger model should pull ahead is long documents: terms that recur across dozens of pages, cross-references, and tables split across pages. We haven't tested that yet.

When to use Fable 5.2 for translation

Use Fable 5.2 forClaude Opus 5.5 is usually enough for
Long documents translated in one pass: contracts, manuals, theses, booksShort, routine texts: emails, product copy, everyday documents
Dense technical or scientific material with ambiguous termsMaterial with a fixed, simple vocabulary
Medical, legal and security documents where a mistaken refusal costs timeHigh-volume jobs where cost per document matters most
Sources in less common languagesCommon language pairs such as English and Chinese

Fable 5.2 costs twice as much per token as Claude Opus 5.5, and its thinking is always on, so check cost per document rather than per token. The best way to decide is to translate one of your own files with both and compare.

❓ FAQ

Is Claude Fable 5.2 good at translation?

In our 10-scenario test it produced accurate translations with consistent terminology, intact code and tables, and corrected OCR errors. Anthropic hasn't published a translation benchmark, so there's no official measure of how much better it is than earlier models.

Is there an official translation benchmark for Claude Fable 5.2?

No. The closest official result is Global MMLU, a multiple-choice knowledge test in 42 languages, where Fable 5.2 scored 96.0% against 94.3% for Claude Opus 5.5. It measures understanding, not translation output.

Does SWE-bench Multilingual show Fable 5.2 is better at multilingual documents?

No. SWE-bench Multilingual covers software engineering problems in nine programming languages. It isn't about human languages or translation.

Claude Fable 5.2 or Claude Opus 5.5 for translation?

Start with Claude Opus 5.5, as Anthropic recommends. It costs half as much per token. Test Fable 5.2 on long, complex documents where consistency across the whole file matters most.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1. Fast mode costs $16 / $80.

Can I use Claude Fable 5.2 in Belin Doc?

Not at the moment. You can translate documents with the models available in Belin Doc and switch between models on the same file to compare results; see our AI model selection guide.

Sources

Checked on November 11, 2026.

  • Anthropic, "Claude Fable 5.2 and Claude Mythos 5.2" — anthropic.com
  • Anthropic, Claude Fable 5.2 / Claude Mythos 5.2 System Card, sections 1.1, 8.1, 8.2 and 8.17
  • Our test material and outputs: 10 scenarios, one run, November 11, 2026

Disclosure: this review is published by Belin Doc, a document translation service. Claude Fable 5.2 isn't currently offered in Belin Doc.

Related Posts