AI Cost Guide

This page explains how AI costs work in Supervertaler for Trados and how to keep them under control. It deliberately avoids quoting exact per-model prices – those change often, and Supervertaler already shows you the real, current cost of every operation in the Reports tab. For exact figures, see Estimates vs actual cost below.

Estimates vs actual cost

The Reports tab shows real billed token counts and cost as reported by your provider’s API – with cache-hit tokens broken out (e.g. 830,000 in (720,000 cached) / 32,000 out · $1.36). When the cost has no ~ prefix, that’s the actual amount your provider will charge. This is the single best place to see what you’re actually spending – it’s live, per-operation, and provider-reported.

The number is cache-aware:

Anthropic (Claude) native + OpenRouter → Claude: real usage from usage.input_tokens + cache_creation_input_tokens + cache_read_input_tokens. Cache reads are billed at 0.1× the input rate (0.05× on Claude Opus 5.5, 0.025× on Claude Fable 5.1), cache writes at 1.25×. See Prompt caching below.
OpenAI: real prompt_tokens and completion_tokens plus prompt_tokens_details.cached_tokens for the auto-cache discount (90–95% off cached input on the current models).
DeepSeek: real prompt_tokens / completion_tokens with auto-cache (90% off cached input).
Gemini 2.5+: real usageMetadata with implicit cache (75% off cached input; 95% off on Gemini 4 Argon, from v18/19.20.203).

For these providers the in-app number is the authoritative billable figure (modulo any account-level credits or monthly minimums you may have).

The chars/4 estimate is still used as a fallback when the provider didn’t return usage info – this affects Ollama (local, free anyway), some provider edge cases, and any response shape we couldn’t parse. In those cases the cost still appears with a ~ prefix to flag it as an estimate.

If you want to cross-check against your provider’s own dashboard:

Provider	Where to look
Anthropic (Claude)	platform.claude.com – Cost
OpenAI (GPT)	platform.openai.com – Usage
Google (Gemini)	console.cloud.google.com – Billing reports
xAI (Grok)	console.x.ai – Usage
Mistral AI	console.mistral.ai – Usage
OpenRouter	openrouter.ai – Activity
DeepSeek	platform.deepseek.com – Usage
Ollama	Free – local execution, no provider console.

The in-app number and the provider dashboard should agree to within rounding for any given run. If you see a meaningful gap, the most likely causes (in order) are: a provider-side credit or volume discount the in-app calculator can’t see; the in-app pricing table being a little behind a recent rate change; or, for fallback (estimate) cases, the chars/4 heuristic over- or under-counting tokens for that particular language and content type.

How costs are calculated

AI providers charge per token – a unit of text roughly equal to ¾ of a word. Costs depend on:

Input tokens – the text you send (source segment, system prompt, terminology context)
Output tokens – the text the model returns (translated segment, proofread text, generated prompt)

Because Supervertaler translates segment by segment, the system prompt and terminology context are included with every segment. For a typical 5,000-word document (~250 segments), the token usage works out roughly like this:

Task	Input tokens	Output tokens
Batch Translate	~125,000	~8,000
AI Proofreader	~140,000	~8,000
AutoPrompt	~10,000	~2,000

These are estimates for a representative document – actual usage varies with segment length, terminology context size, and prompt complexity. Token counts like these are fairly stable; what changes is the price per token, which is why this guide points you to live figures rather than quoting them. Most of those input tokens are the same in every request, which is what prompt caching makes cheaper.

Prompt caching

Most of what Supervertaler sends with each request is the same every time: the instructions, your custom prompt, your SuperMemory context, the terminology and the document context. Only the segments change. Prompt caching lets the AI provider keep that shared part after the first request and charge much less for it in the requests that follow.

Where it’s used. Batch Translate, the AI Proofreader, Translate active segment (Alt+T) and, from v18/19.20.205, chat in the AI Assistant. All of them keep the shared part of the prompt exactly the same from one request to the next, so it can be cached. AutoPrompt and QuickLauncher prompts are one-off requests and are sent without caching.

In the AI Assistant chat (from v18/19.20.205), every message you send goes to the AI together with the instructions, your SuperMemory context, the document and the conversation so far. From the second message on, all of that is read from the cache, and only your new message and the details of the active segment – its position, the segments around it, its TM matches and its terms – are paid in full. When you move to another segment, those details and the conversation are sent again in full, but they are short; the document part is still read from the cache. When the AI uses a tool to answer, each step reads the one before it from the cache too.

With Claude (directly, or through OpenRouter), Supervertaler marks the shared part of the prompt for Claude’s cache:

The first request pays 25% extra on that part, to store it.
Every request after that reads it back for a tenth of the normal input price, or less: a twentieth on Claude Opus 5.5 and a fortieth on Claude Fable 5.1.
The cache lasts five minutes, and every request that reads from it starts the five minutes again. As long as you keep working, it stays available.

A prompt shorter than Claude’s minimum is simply not cached, and costs nothing extra.

OpenAI, DeepSeek and Gemini (2.5 and later) cache automatically, with no marker needed. Because Supervertaler keeps the shared part of the prompt the same, they give their cache discount too – see Estimates vs actual cost for the rates.

Why the five-minute cache. Claude also offers a one-hour cache, but storing a prompt in it costs twice the normal input price instead of 1.25×. In a batch run the requests follow each other within seconds, and when you translate segment by segment each new segment restarts the five minutes, so the shorter, cheaper cache is enough.

What makes it start over. The cache only helps when the shared part is exactly the same. The next request pays the full price again (plus the 25% with Claude) after:

a pause of more than five minutes;
a change of model, prompt or AI settings, or a switch to another SuperMemory bank;
moving to another document, or starting a new batch run over different segments (each run sends only the terminology its own segments need).

Seeing it work. In the Reports tab, the number of tokens read from the cache appears next to the input count, e.g. 830,000 in (720,000 cached) / 32,000 out · $1.36, and the cost already includes the discount. The Token Usage & Costs log keeps cache reads and cache writes in separate columns.

How much will it cost?

There’s a wide spread between models. As a rough mental model:

Local models (Ollama) are free – they run on your own computer, with no API charges at all. The trade-off is that quality depends on your hardware, and they’re generally less capable than cloud-hosted models. If you have a computer with 8+ GB of RAM, TranslateGemma 12B delivers surprisingly good results for free.
Budget cloud models – the cheapest tier from each provider (e.g. GPT-6 Luna, Gemini 3.5 Flash-Lite, Mistral Small, DeepSeek V4 Flash) – typically cost a small fraction of a cent per segment. They’re excellent for routine, high-volume translation.
Flagship models – Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and the like – can run roughly 10–100× the price of the budget tier. Reserve them for specialised content where the quality difference earns its keep.

To see what a model actually costs for your work, run one operation and check the Reports tab – it shows the real billed cost. For OpenRouter, expect the underlying provider’s rate plus a small platform fee.

Our recommendation

For budget-conscious batch work, GPT-6 Luna or Gemini 3.5 Flash-Lite offer excellent quality at a fraction of the price. For the absolute highest quality on specialised content, Claude Opus 5.5 or GPT-6 Astra are worth the premium.

Token pricing

Supervertaler’s in-app cost figures come from a built-in per-token pricing table. Because provider prices change regularly, that table is occasionally a little behind a recent rate change – the Reports tab’s provider-reported figures are always the authoritative ones. For the definitive current rates, check the provider’s own pricing page:

OpenAI · Anthropic · Google Gemini · xAI · Mistral · DeepSeek · OpenRouter

Tips for managing costs

Start with a budget model – GPT-6 Luna, Gemini 3.5 Flash-Lite, or Mistral Small are excellent for routine translation at a fraction of the cost of a flagship.
Use premium models selectively – reserve GPT-6 Astra, Claude Opus 5.5, or Gemini 4 Argon for specialised content (legal, medical, patents) where the quality difference justifies the cost.
Try Ollama for zero cost – if you have a computer with 8+ GB of RAM, TranslateGemma 12B delivers surprisingly good results for free.
Check your usage – the Reports tab lists every AI call live with its token count and cost; the Token Usage & Costs report totals your spend over time (by project, client, model or month) and exports it to CSV/Excel; and your provider’s own console (see the Estimates vs actual cost table above) shows the authoritative billable figure.
Set a monthly budget – give Supervertaler a soft monthly limit (Settings → AI Settings) and it will warn you before a batch once you’ve reached it. See Token Usage & Costs.
Keep going when translating segment by segment – from v18/19.20.198, Translate active segment (Alt+T) sends the same shared prompt every time, so a provider that caches prompts charges about a tenth for it from the second segment on, if each segment follows within five minutes. A pause longer than that means the next segment pays full price again. See Prompt caching and Translate Active Segment.

Built-in cost protection

Supervertaler includes several safeguards to help you avoid unexpected costs:

QuickLauncher prompts are standalone

When you run a prompt from the QuickLauncher menu (Alt+Q), only the prompt itself is sent to the AI – not the chat history. This means a simple terminology query costs only what it needs to, even if you have a long conversation in the chat window.

Chat token budget

Regular chat messages include recent conversation history so the AI can follow your discussion: your last 10 to 13 messages. From v18/19.20.205 the oldest ones are dropped two exchanges at a time rather than one with every message, so the conversation keeps the same beginning for several messages and can be read from the cache. Supervertaler also trims older messages when the history grows too large (~50,000 tokens). This prevents costs from spiralling when previous messages contained large context blocks (e.g. full document content).

Cost warning

If a request is estimated to cost more than $0.50 in input tokens, a confirmation dialogue appears showing the estimated token count and cost. You can cancel before the expensive request is sent.

Choosing the right model

For everyday work – chat queries, terminology questions, QuickLauncher prompts – use GPT-6 Luna or another budget model. Reserve premium models like GPT-6 Astra or Claude Opus 5.5 for AutoPrompt and complex tasks where the quality difference justifies the cost.