Every claim on this page is a measurement you can re-run, made on the complete published texts — not a marketing line. The method is the product, so we show the numbers.
If you download a free Herodotus or Josephus today, this is who actually wrote your English:
| Author | The standard free translation | Age today |
|---|---|---|
| Josephus | William Whiston | 289 years |
| Philo | C. D. Yonge | 172 years |
| Plato | Benjamin Jowett | 155 years |
| Herodotus | G. C. Macaulay | 136 years |
| Homer | Samuel Butler | 128 years |
| Hesiod | H. G. Evelyn-White | 112 years |
Seven translators across two centuries, each with his own century's English. Good modern translations exist — but they are sold per volume, by different translators with different habits, and never as one voice across the whole library.
We counted the archaic tokens (thee, thou, hath, unto, spake, whosoever…) per 10,000 words in the standard free translation of each author, and in ours. Full texts, same tokenizer, same list:
Archaic tokens per 10,000 words, full published texts. Sentence length drops too where the source allows it (Philo: 45.2 → 39.5 words per sentence; Herodotus: 42.8 → 32.8) — but we never chop the author's rhetoric into fragments to fake readability.
The obvious worry about AI translation is that it might quietly reproduce an existing one. So we test for exactly that, mechanically: we scan every published book for any run of 8 or more consecutive identical words shared with the prior translations of the same text — and we publish the calibration.
The human baseline experiment. How much do two independent human translators of the same Greek overlap by pure convergence? We compared two professional translations of Herodotus Book 1 against each other: 18 shared 8-word runs (longest: 10 words). Our Book 1 against the modern one: 3 runs (longest: 10). Our text overlaps existing translations six times less than two humans overlap each other.
Every book on the shelf shipped at zero 8-word runs against the public-domain translation of the same passages, and the most-quoted passages are additionally spot-checked against the modern copyrighted translations. Whole-corpus 3-gram overlap with the classic versions runs 4.7–8% — the same range as two unrelated prose translations of the same story. Details and per-book numbers: the proof page.
The hardest question about AI translation is accuracy — and AI checking AI can share blind spots. So we calibrated the checkers: we drew 60 random passages (about 11,500 words) from every book in the library, secretly mixed in 8 sabotaged passages with deliberately corrupted meanings (flipped negations, changed numbers, swapped agents), and had independent AI graders judge all 68 blind, sentence-by-sentence against the Greek and Latin.
Results. The graders caught 7 of the 8 planted corruptions (the one miss was a negation buried in the Sophist's most abstract dialectic). On the real 60: 49 fully accurate, 9 minor nuance notes, 2 genuine errors — both in one book (Seneca), both traced to one root cause (paragraph-level truncation that letter-level checks missed), both repaired the same day along with 10 sibling defects the follow-up sweep found. The Greek corpus sampled zero major errors, and a corpus-wide paragraph-level sweep confirmed zero suspect sections across all 2.9 million Greek-derived words.
We then scaled that method to the whole library. Every one of the nine works was audited against its original Greek or Latin: 620 passages sampled across every book, each paired with its actual source text, each graded by an independent AI philologist for accuracy only — and every flagged problem then attacked by three adversarial reviewers instructed to refute it. A claim only counted as an error if the source clearly contradicted the English under that hostile review. (This required first building the one missing source: a complete Moralia Greek text, assembled from the Perseus Digital Library and validated essay-by-essay — now offered on the open-data page.)
Results. Of 443 source-comparable passages, 426 were clean — 96.2%. The 17 confirmed errors (subtle referent slips, one flipped negation, one duplicated sentence, a mule rendered as an ox) were each corrected with a minimal fix, and every fix was independently re-verified against the source before publishing. A parallel mechanical sweep of all 3.6 million words found zero truncations. The corrected error list and the audit tooling are preserved in the project archive; the errors clustered exactly where philology predicts — in the densest authors (Plato, Philo) — which is itself evidence the audit measured something real.
Silent omission is the most dangerous translation defect — a missing paragraph reads perfectly and no one notices without the original. Every batch we publish passes mechanical completeness gates (aligned section counts against the Greek; word-ratio screens per chapter), and independent AI referees re-read the source behind samples of every book hunting for gaps. When our own audit found six omissions in our first editions, we published the repairs the same day — the process is the warranty.
This shelf is about 3.1 million words of translation. A professional classicist translates roughly a million words in a working lifetime alongside their other duties — which is why no single human has ever translated Homer, Hesiod, Herodotus, Plato, Plutarch's essays, Josephus, Philo and Seneca. It has always been a patchwork of centuries and voices. One voice matters for readers (you learn one register, then simply read) and for scholarship: when every author speaks the same English, differences you notice between books are differences in the books — not artifacts of Whiston's 1737 habits colliding with Jowett's 1871 ones.
Every book in this library was produced under the direction of Claude Fable 5, Anthropic's most capable model — the strongest AI available to us, holding the final pen on every page. Fable 5 coordinates the whole pipeline, arbitrates every disputed finding, rewrites every flagged passage, performs the independence rewrites, and translates the hardest material itself; Claude Sonnet fleets carry the bulk translation under its direction, and Claude Opus independently referees samples of every book against the Greek. In our head-to-head trials on the same Greek, Fable 5 was the only model that never silently dropped a line — that is why it holds the final pen. The full division of labor is in the colophon.
We metered the whole production run. Translating, checking, auditing, and repairing all ~3.1 million words consumed roughly 110–160 million AI tokens across the three-model pipeline. At the raw July-2026 API list prices (Claude Sonnet $3/$15, Opus $5/$25, Fable 5 $10/$50 per million tokens in/out), that is approximately:
| Stage | Est. raw API cost |
|---|---|
| Bulk translation (Sonnet fleets, ~1,550 batches) | ~$650 |
| Referees, audits, independence rewrites | ~$350–450 |
| Claude Fable 5 — coordination, arbitration, hard passages | ~$400–800 |
| Total | ~$1,400–$2,000 |
For comparison: professional literary translation runs $0.15–$0.25 per word. A human retranslation of this shelf would cost roughly $465,000–$775,000 and several careers of specialist time. Frontier AI collapsed that by two-and-a-half orders of magnitude — that is why this library can be given away, and why nothing like it existed before 2026.
Estimate from metered per-batch token usage (measured mean: ~52,000 tokens per 1,500-word batch) extrapolated across all production and QC fleets; orchestration estimated conservatively. Actual production ran on subscription plans; the figures show what the compute would cost at pay-as-you-go API prices.
These translations exist partly for scholarship: one consistent voice across the whole corpus removes the translator variable from any English-side comparison (motif searches, register analysis, cross-author studies), and the entire production method — sources, checks, error rates, repairs — is published on this site. The original Greek and Latin always remain the authority; treat these as you would any translation: verify load-bearing readings against the source text.
Suggested citation:
Herodotus: The Histories. The Plainspoken Classics, first edition. AI translation directed by Claude Fable 5 (Anthropic), Scriptorium Press, 2026. openscriptorium.com/books/herodotus. Accessed [date].
Substitute the work and slug as appropriate. Editions are versioned: the text you cite is the "First Edition (2026)" shown on each volume, and corrections are logged publicly in the site's repository history. For archival permanence, the TXT and EPUB downloads let you deposit the exact text you used alongside your data. Questions or corrections: the contact address in the footer — errata reports are welcomed and credited.
Every book is free to read right now, cover to cover, with the first chapters as read-along samples. If the old translations serve you better, they are a click away and always will be — we just think that after one chapter you won't go back.
Browse the library — every book free See the raw numbers Read the methodAll metrics measured on the complete published texts (07/2026). Archaism list: 28 tokens (thee/thou/thy/hath/doth/unto/spake/whosoever…), case-insensitive, per 10,000 words. Independence scans: normalized 8-gram set intersection. Errors are possible in any translation; the Greek and Latin remain the authority, and every edition carries its AI-translation disclosure.