Token Usage & Costs

Supervertaler Workbench keeps a persistent log of the AI tokens and cost of every operation, plus a built-in Usage & Costs report to total and export it. Use it to answer “how much did this project cost?”, “how many tokens did we use this month?”, or “what should I bill this client for AI?” – across every provider, including local and custom models.

You can also see what one project has cost so far, get an estimate before a batch run, and show costs in euros.

The log uses the same format as the Supervertaler for Trados plugin, so if you use both, their logs merge into a single analysis.

The usage log file

Every AI call appends one line to a monthly file:

…\Supervertaler\workbench\usage\usage-2026-06.jsonl

It is JSONL (one JSON object per line) – open it in Excel, or parse it with a script. A line looks like:

{"ts":"2026-06-18T16:07:03Z","product":"workbench","task":"BatchTranslate",
 "provider":"claude","model":"claude-sonnet-4-6","project":"My Patent Job",
 "file":"","src_lang":"English","tgt_lang":"Dutch","in_regular":654,
 "in_cache_read":0,"in_cache_write":27843,"out":808,"source":"actual",
 "cost_usd":0.11849325,"cost_known":true,"duration_s":24.9,"ok":true}

It is on by default. Turn it off in Settings → AI Settings → AI Cost Monitoring.

The Usage & Costs report

Open Tools → 💰 Token Usage & Costs…. The window totals your usage and lets you slice it:

Range – This month, Last 3 months, This year, or All time.
Group by – Project, Client, Model, Provider, Task, Day or Month.

Each row shows calls, input/output tokens, cost, and a % actual column (the share backed by provider-reported figures rather than estimates). The footer shows the range total and your month-to-date spend against your budget.

Showing costs in euros

The report has a Show costs in: switch with USD and EUR. Choose EUR and a 1 USD = box appears next to it with the exchange rate. Supervertaler doesn’t look the rate up online, so set it to your bank’s or card’s rate.

Your choice is remembered, and it also applies to Project Information and the Batch Translate estimate. The price list and the exported ledger stay in US dollars, as in Supervertaler for Trados.

Exporting

Export CSV… and Export Excel… write the detailed ledger (one row per call) for the selected range – ready for invoicing or analysis.

The cost of one project

Project → 📋 Project Info… opens the Project Information window. Its AI & Prompts section shows:

AI Model – the model the project is being translated with, and its provider. For a custom OpenAI-compatible endpoint it also names the active profile. It always matches the model shown in the status bar.
AI cost so far – what the project has cost in AI: the total, the number of AI calls, and the tokens in and out. If some calls used a model that isn’t in the price list, it says how many.
Estimated cost to AI-translate the rest – an estimate for the segments that are still empty, with your current model. A local (Ollama) model shows as free; a model that isn’t in the price list shows as unknown.

The figures come from the usage log and are matched on the project name. Calls made while the log was switched off aren’t counted.

Every AI call made while a project is open counts towards that project: batch translation, but also Ctrl+T, the Chat assistant and AutoTagger. (Before v1.10.372 only batch translation recorded its project, so the others were listed under “(none)” when you grouped the report by Project.)

An estimate before you start

The Batch Translate dialog shows what the run is likely to cost before you click Start Translation, for example:

💰 Estimated cost: ~$0.42 · ~120,000 tokens in / ~35,000 out in 12 call(s)

The estimate is worked out the way the batch will actually be sent: with your current prompt and glossary, your batch size, and the same price list as the usage log. When your prompt is long enough for the provider to cache it, batches 2 onwards are priced at the cheaper cached rate, as they will be billed. Token counts are estimated at about 4 characters per token, so treat the figure as a guide; the real cost of each call is logged as usual. See Batch Translation for details.

Settings & budget

Settings → AI Settings → AI Cost Monitoring has:

Keep a persistent token-usage log – the on/off switch.
Monthly budget (USD) – a soft monthly limit (cents allowed; 0 disables). It is always entered in US dollars, even when you show costs in euros. Once this month’s logged spend reaches it, starting a batch translation shows a warn-and-continue prompt. It is advisory and never blocks.

Pricing custom / self-hosted models

Costs come from a single price list, pricing.json, shared with the Trados plugin. To price a custom or self-hosted model – or override any rate for both products at once – copy the bundled modules/pricing.json to …\Supervertaler\pricing.json and add an entry keyed by the exact model id:

{ "models": { "my-university-llama": { "input": 0.0, "output": 0.0 } } }

Until a rate is set, a custom model’s tokens are still logged, with the cost marked unknown rather than guessed. The Batch Translate estimate and Project Information say “unknown” for such a model too, rather than pretending it is free. Local models (Ollama) are priced at 0.

How accurate are the figures?

Every record is flagged actual or estimated in its source field – and the difference matters.

actual (the usual case): the token counts are the exact numbers the provider’s API reported for that call – the same numbers it bills you against. These are as accurate as it gets. This covers OpenAI, Claude, Gemini, Mistral, DeepSeek and OpenRouter, including the cached-token breakdown.
estimated: a fallback used only when the provider returned no usage data – currently local models (Ollama) and the occasional unparseable response. The estimate is a simple characters ÷ 4 heuristic. It’s reasonable for English but can be well off for other content: scripts such as Chinese, Japanese, Korean, Arabic and Cyrillic pack a different number of characters per token, so the estimate often under-counts them. Ollama is free, so for local models only the token count is approximate – there is no cost involved.

For cost, an actual record’s figure is the exact token count × the per-model rate from pricing.json, with cached tokens priced at each provider’s cache discount (e.g. Claude cache reads at 10% of the input rate and cache writes at 125%, OpenAI cache reads at 50%, Gemini 2.5+/3 at 25%). This matches how the Supervertaler for Trados plugin computes cost, so the two products agree. The one thing that can make it differ slightly from your provider’s bill:

The price list can lag a provider’s recent rate change. The model id and the token counts are still exact – only the unit price might be a little behind.

For the definitive bill, your provider’s own usage dashboard is authoritative. The ledger is built for tracking trends, attributing usage to projects and clients, and capacity planning – where the exact, provider-reported token counts are precisely what you want.