Repository navigation
Choosing a Model
There's no single right answer, but there are sensible defaults. This page covers where a stronger model pays off and where a cheap one is fine.
If you're new, connect one model and let it do every step. Polygluttony marks a sensible pick for each provider:
| Provider | Recommended | Also suggested |
|---|---|---|
| Anthropic | Claude Sonnet 5.5 | Claude Opus 5.5 (best quality), Claude Haiku 5.5 (fastest & cheapest) |
| OpenAI | GPT-5.5 | GPT-5.4 mini (fastest & cheapest), GPT-4.1 |
| Gemini 3.1 Pro | Gemini 3 Flash (fastest), Gemini 2.5 Pro, Gemini 2.5 Flash |
Translate a couple of episodes and read them. If you're happy, stop here.
Each step does a different amount of work, so each benefits differently from a stronger model:
| Step | How much it sends | What matters | Suggestion |
|---|---|---|---|
| Find names | Every line of the episodes it reads; later, only new episodes | Spotting names and terms reliably | A good mid-range model |
| Tidy up | One small request per category | Judgment: which spelling is right, what's too generic | A strong model; it costs little here |
| Look up names online | One request each time you run it | Finding what fans call things | Any model that can search |
| Translate | Every line, plus context | Quality: this is what your viewers read | The best model you can afford |
| Verify | A sample of lines | Spotting lines that are completely off | A cheap, fast model is fine |
The app gives the same hint: "e.g. a stronger model to tidy up, a cheaper one to verify".
An example that works well: Claude Sonnet 5.5 for Find names, Translate and Look up names online, Claude Opus 5.5 for Tidy up, Claude Haiku 5.5 for Verify.
See Connecting a model for how to split them.
For Translate, higher thinking helps most on dense dialogue: wordplay, cultivation jargon, lines that only make sense with what came before. It also makes each batch slower and uses more output tokens. High is a good default for translating; Medium is plenty for Tidy up and Verify.
Polygluttony is free. Your provider bills you for tokens, the text going in and coming out. Roughly:
- Cost grows with the number of lines. Each request also repeats the prompt, the matching glossary terms and a few lines of context.
- Thinking adds output tokens, often more than the translation itself.
- On Anthropic, the prompt is cached between requests, so you don't pay full price for it every time.
For scale: one run with Claude Sonnet 5.5 on High thinking used about 12,000 input tokens (part of them cached) and 10,000 output tokens for each 200-line batch, at about 45 seconds a batch. To see your own numbers, open Logs at the bottom of the Episodes page; every request lists its tokens. Your provider's dashboard has the exact bill.
Ollama and LM Studio cost nothing per line, and nothing leaves your computer. The trade-offs:
- Speed depends on your hardware. By default, Polygluttony sends a local server one request at a time.
- Quality depends heavily on the model. Smaller models get the line format wrong more often, which means more retries, more lines left untranslated, and more episodes that need a look.
- They can't look up names online, and have no Thinking control.
A local model works well for a step that tolerates mistakes, like Verify, while a cloud model translates.
If your provider account has low rate limits, lower Requests at once in the step's Tuning. Polygluttony backs off by itself when a provider says it's sending too much (an episode then shows "Waiting — Anthropic asked us to slow down · 18s"), but starting lower avoids the wasted requests.