What powers Bestsella
Bestsella shows quality tiers in the product — Draft, Standard, Premium — and discloses the exact model routing here. This page renders directly from the same configuration the pipeline uses, so it cannot drift from reality. Every routing change is checked against a published regression gate and recorded in the changelog below; a change made ahead of its gate says so there, with the result added once the gate has run.
Current routing
| Task | Draft(≈2 cr / 1k words) | Standard(≈5 cr / 1k words) | Premium(≈8 cr / 1k words) |
|---|---|---|---|
| Section & chapter prose | Claude Haiku 4.5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Outlines | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Titles | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Descriptions & listing copy | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Author bios | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Author & reader personas | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Inline editing agents | Claude Haiku 4.5 | Claude Sonnet 5.5 | Claude Sonnet 5.5 |
| Worldbuilding, characters & plot tools | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Market research extraction | Claude Haiku 4.5 | Claude Haiku 4.5 | Claude Sonnet 5.5 |
| Compliance analysis | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Sonnet 5.5 |
| Quality reviews | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Launch emails & materials | Claude Sonnet 5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Manuscript import analysis | Claude Haiku 4.5 | Claude Sonnet 5.5 | Claude Opus 5.5 |
| Writing assistant chat | Claude Haiku 4.5 | Claude Sonnet 5.5 | Claude Sonnet 5.5 |
| Cover lettering match | Claude Sonnet 5.5 | Claude Sonnet 5.5 | Claude Sonnet 5.5 |
| Cover & diagram images | gemini-3-pro-image (all tiers — image generation sits outside the quality tiers) | ||
About AI watermarks
Every Claude model Bestsella routes to — in every tier — produces text that carries an invisible, model-level watermark. Anthropic introduced it in August 2026 to meet EU transparency rules. It applies worldwide, and neither Bestsella nor you can switch it off.
Nothing is added to your text. The mark is a statistical pattern in word choices that only a holder of Anthropic's key can test for. It carries no information about you, your account, or your project. It may survive light editing; a full rewrite in your own words removes it — at which point the text is no longer AI-generated in any meaningful sense.
What this means for you: Amazon KDP already requires you to disclose AI-generated text, images, and translations at upload, whether or not you edited them afterward. Manuscripts drafted in Bestsella are AI-generated under that definition. Disclose them. The disclosure is made to Amazon during upload and is not shown to readers.
Bestsella will never offer a feature designed to hide, weaken, or remove this mark, and never a tool to test whether your writing "reads as AI." We build for authors who publish honestly.
Deep operations
Whole-book operations are in dogfooding and will be documented here before launch, including their data retention terms.
Routing changelog
2026-10-07
New task: reading the title lettering on a book's front cover (the nearest bundled typeface, whether it is set in capitals, and its colours) so the spine title can match it, when the author asks. One call that reads the cover image and returns structured data, on Claude Sonnet 5.5 in every tier, Draft included.
Gate: Does not apply: the validator regression gate scores written sections, and this task writes no prose. No earlier routing existed for it, so no model was replaced. Added by owner decision. Request settings match the other structured-extraction tasks on Sonnet 5.5: no sampling parameters, and the lowest thinking setting it accepts.
2026-09-28
Standard moved from Claude Sonnet 5 to Claude Sonnet 5.5, released the same day at the same price, on every task where Standard ran on Sonnet 5 (market research extraction stays on Claude Haiku 4.5). Premium's tasks that ran on Sonnet 5 (the writing assistant, inline edits, market extraction and judgment) moved with it, so Premium never runs an older model than Standard. Draft is unchanged.
Gate: Switched ahead of the validator regression gate, which then passed (2026-09-28): the same 5 benchmark sections (3 non-fiction, 2 fiction) written on Sonnet 5 and Sonnet 5.5 from the same brief and context, then blind-scored on the 10-dimension base rubric by 3 AI judge runs per section (Claude Opus 5.5). Sonnet 5.5 scored higher on 4 of 5 (composite 6.94 vs 6.11), was preferred in 14 of 15 judgements and regressed on no dimension, at about the same total cost. Request settings follow Anthropic's migration notes: no sampling parameters; where thinking used to be turned off it now uses the lowest setting Sonnet 5.5 accepts.
2026-09-28
Premium moved from Claude Opus 5 to Claude Opus 5.5, Anthropic's recommended model for most work, on every task where Premium ran on Opus 5. Opus 5.5 costs about 20% less per token than Opus 5.
Gate: Switched ahead of the validator regression gate, which did not pass (2026-09-28): on the same 5 benchmark sections, blind-scored the same way (3 AI judge runs per section, Claude Opus 5.5), Opus 5.5 scored consistently lower on prose, pacing, voice and section openings, and Opus 5 was preferred in all 15 judgements; the overall composite stayed inside the noise band (6.82 vs 7.10). Opus 5.5 cost about 5% less in total across the 5 sections: it writes more, which offsets most of its 20% lower per-token price. Kept on Premium by owner decision; it will be re-tested before any further Premium change. Request settings follow Anthropic's migration guide: no sampling parameters, thinking never disabled, effort set on every request (high, the level Premium ran at on Opus 5; low for extraction tasks).
2026-08-27
Routing moved to the current model generation: Standard prose to Claude Sonnet 5, Premium prose to Claude Opus 5, Draft prose to Claude Haiku 4.5 (declared iteration-grade on its card).
Gate: Validator regression gate passed: 5 blind-scored section pairs (3 non-fiction, 2 fiction), old vs new routing — new routing preferred 4/5, composite 7.26 vs 6.98, no dimension regressed consistently; measured per-call cost equal or lower despite ~1.44x tokenizer inflation.
2026-08-27
Tier system launched: Draft / Standard / Premium routing profiles replace the per-model picker. Routing itself unchanged at launch (mirrored prior production models).
Gate: Behavior-neutral by construction — the model-generation flip ran separately behind the regression gate above.
Model catalog and prices last verified 2026-09-28. Pricing basis: Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5.5, Claude Sonnet 5, Claude Sonnet 4.6, Claude Opus 4.8, Claude Haiku 4.5.