Translation Cost & Memory

This page is the audit + optimisation surface for paid translation. It tracks every translation job’s cost, surfaces hit-rate of the translation memory (TM), and lets you prune the TM when it gets stale. The goal is simple: pay once per unique cue, not once per re-run.

Where it lives

Settings → Translation → Cost & Memory — split into two tabs that share the page:

Tab	Content
Cost	Per-job cost log + aggregate by backend, day, and series.
Memory	Translation memory entries with hit count, last-used timestamp, source + target.

Cost tab

Every completed translation job writes a row to translation_events:

Field	Effect
Job ID	Unique identifier; click to open the produced file.
Series / Movie	Source media.
Backend	The provider that did the work.
Source language → Target	Language pair.
Tokens / chars in	Input volume sent to the backend.
Tokens / chars out	Output volume returned.
Latency	Wall-clock time from queue to finish, in seconds.
Cost (USD)	Computed from in/out volumes × backend’s configured price.
TM hit rate	Fraction of cues served from translation memory rather than the backend.

Aggregations

The header shows running totals:

Tile	Period
Today	Jobs in the current server-day.
Last 7 days	Rolling week.
Last 30 days	Rolling month.
All time	Since the database was created.

Each tile breaks down per backend so you can see where spend lives. Click a tile to expand into a per-backend daily chart.

Per-backend price

Cost is computed at job time from the configured price under Settings → Translation → Backends → <backend> → Price. Changing the price doesn’t retroactively rewrite history — but the new price applies to every new job. For OpenAI / Anthropic / Gemini, prices are per 1M input + output tokens; for DeepL / Azure, per 1M characters.

Caution

Update the price when your backend bills you a new rate. Sublarr can’t read upstream pricing pages — what you enter is what cost columns show.

Memory tab

Translation memory caches (source language, source text, target language, backend) → translated text. On every translation job, Sublarr checks the cache first; only cache misses hit the backend.

Why it matters

Scenario	Without TM	With TM
Re-translate same episode after editing a cue	Re-translate the entire file	Hit cache for unchanged cues, only pay for the edited one
Translate a recap episode	Pay full cost for all the recapped lines	Cache hit — recap lines already translated for prior episodes
Same cue across multiple shows (e.g. opening credits)	Pay each time	Pay once

Memory entries

Column	Effect
Source text	Cue text exactly as it appeared.
Target text	Translated rendering.
Source / target language	Language pair.
Backend	Backend that produced this entry — the cache is keyed per backend so quality stays consistent.
Hits	How many times this entry was served from cache.
Last used	Timestamp of the most-recent hit.
Created	Original translation timestamp.

Filtering and search

Filter	Effect
Search	Substring match on source or target text.
Backend	Restrict to one backend.
Language pair	en → de, ja → en, etc.
Min hits	Only entries used at least N times — finds your “load-bearing” cache.
Last used	”Older than 90 days” — finds candidates for pruning.

Pruning

When the TM grows too large or contains entries from a backend you no longer use:

Action	Effect
Delete selected	Remove specific rows.
Delete by filter	Delete every row matching the current filter. Confirm prompt shows count.
Delete by backend	Drop everything from a specific backend (after a quality regression, e.g.).
Reset	Wipe the entire TM. Future jobs start with cold cache.

Caution

A wipe means every cue gets retranslated and paid for again. Reserve it for major quality changes (new prompt template, switched primary backend) where re-running everything is cheaper than letting old quality leak through.

Hit-rate optimisation

The TM hit rate per job is a quick proxy for how much money you’re saving:

Hit rate	Reading
0%	Cold start or unique content.
10–30%	Typical for fresh anime — a lot of unique dialogue.
40–60%	Recap episodes, sequels reusing OP/ED, sitcoms with catchphrases.
70%+	Re-translation after an edit; or repetitive content (clip shows, training videos).

If hit rate is consistently below your expectation:

1. Confirm the same backend is used (cache is keyed per backend).
2. Confirm the source language hasn’t been re-detected (a swap from en to und would miss the cache).
3. Ensure the source file hasn’t been re-encoded with subtle whitespace differences — the TM is exact-match on text.

Memory across backends

Each backend has its own cache because translation quality differs. Switching the primary backend means:

Effect	What happens
First job after switch	Cache miss for everything; full cost on the new backend.
Old backend’s cache	Still present; usable if you switch back.
Hybrid runs	Mid-job backend swap (e.g. Ollama failed, fall back to Claude) creates two separate cache entries per cue.

Use Settings → Translation → Memory → Backend filter to inspect each backend’s cache independently.

Export for analysis

Both cost rows and memory entries can be exported as CSV from the page header. Useful for:

• Monthly cost reconciliation against your backend’s invoice.
• Sharing with auditors.
• Bulk editing the TM externally and re-importing.