Choose OpenAI, Gemini or Claude for AI editing

What the AI editing pass actually does

RoughCut removes silence entirely on your Mac, with no AI and no key. The AI pass is a separate, optional stage that reads the transcript and proposes removing spans it judges to be retakes, reworded retakes, false starts, tangents, filler, or noise. Every proposal lands in the review list with a category and a confidence score, and nothing is removed from your source file.

What is sent: transcript text and word timing context. What is never sent: your video, your audio, or your files. The one exception in the whole product is Pro Cloud transcription, which is a different feature with its own per-job opt-in.

When you need no key

Turn Bad-take removal off in the New Rough Cut screen. RoughCut then runs a silence-only cut using on-device analysis alone. This needs no provider, no key, no billing, and no network. The review list shows LLM editor: skipped — silence-only cut.

This is a genuinely useful mode, not a crippled fallback — for clean, well-rehearsed recordings, silence trimming alone often gets you most of the way.

Comparing the three providers

RoughCut does not recommend one provider over another. OpenAI is the default for new installations, so the welcome tour's key page and the dated example in Help → API Costs use it; Gemini and Claude are chosen in Settings. The LLM provider setting's own help text is the rule: OpenAI is the default for new installations. gemini is Google Gemini and anthropic is Claude. Each provider uses its own API key and bills you directly. Existing provider choices are preserved; long transcripts are split into chunks and merged back on every provider. If you updated from an earlier version with Gemini selected, nothing changes until you change it.

If you plan to use Pro Cloud transcription, you need an OpenAI key for that specifically, because Cloud transcription is an OpenAI feature. You can still run the editing pass on Gemini or Claude — the transcription source and the editing provider are two independent settings.

Three buttons that save different things

RoughCut keeps keys, the provider choice and the model choice in different places, and each has its own button. Knowing which is which avoids the most common "I saved it but nothing changed" confusion.

The app says the same thing in the model picker: Use Model updates this provider's model field. Save Configuration in Settings to apply it to future jobs.

Choose a reasoning level

Some OpenAI and Claude models let you choose how long the model thinks before it answers. In Choose Model & Estimate… next to OpenAI model or Claude model, a Reasoning level picker lists the levels the selected model supports.

The first entry is the default, and its label tells you what it means for the selected model: Default (medium) for GPT-5.6 Sol, GPT-5.6, GPT-6 Astra and Claude Opus 5.5, Default (high) for Claude Sonnet 5.5 and Claude Fable 5.1, and Provider default where RoughCut does not know the provider's own default. On OpenAI, Default keeps RoughCut's tested setting: medium on Sol, GPT-5.6 and Astra, and nothing sent for other model IDs. On Claude, Default sends no level, so the model's own default applies.

The level words stay as the providers write them, in both languages. Under the picker the app states the trade-off: Higher levels think longer: slower and more expensive, and not always better. Reasoning tokens are billed as output. RoughCut makes no claim that a higher level cuts better — compare on a recording you know before you change it for real work.

What else to know:

• Use Model stores the model and the level together, and Save Configuration applies them. The model row on Settings → Basic then reads, for example, gpt-5.6-sol · reasoning default.
• The level is used by the first AI pass and by Second Look.
• A higher level allows a longer wait: at least 900 seconds at high, 1800 at xhigh and 2700 at max. A longer timeout you configured yourself is kept.
• If a model does not offer the level you picked, Use Model stays disabled and the picker says This model does not support the selected reasoning level. Pick another level. A job that still reaches the AI stage with such a level stops before any provider call, with a message that lists the supported levels. See Error message reference.

Estimate the cost before you commit

Next to each model field, Choose Model & Estimate… opens a picker that lists the models your saved key can reach (Refresh Models) and a calculator with editable assumptions: Recording duration (minutes), Estimated edit passes, input and output tokens per 30 minutes, and Token prices (USD per million). Every one of those is a planning assumption you can change — the result is a planning estimate, not a price, a quote, or a spending cap. The picker says so itself: Illustrative estimate, not a measured average, quote or spending cap. Duration alone does not determine token usage. Without custom rates it uses Published standard text rates stored in this version of RoughCut. Prices may change; verify the provider's pricing page.

Help → API Costs gives the shorter version: a dated example for the default model, a reminder that reasoning tokens count as output, that longer transcripts, extra editor samples, Second Look and retries add cost, and links to the current OpenAI, Gemini and Claude rates. The bill is whatever your provider's dashboard shows.

Play from 1:50 · Check the model and estimate4 min · English captions

What it costs

No provider is free, and RoughCut deliberately does not publish a per-video figure.

• OpenAI bills API usage against a funded API balance — prepaid credits or a payment method on the Platform organisation. A ChatGPT subscription does not fund API usage, and neither does buying RoughCut or RoughCut Pro.
• The default Gemini model is on Google's paid tier, so billing must be enabled on the Google project behind your key. A Gemini app subscription does not fund API usage either.
• Anthropic bills Claude API usage separately, through billing set up in the Claude Console.
• A higher reasoning level costs more, because reasoning tokens are billed as output.
• On-device transcription costs nothing extra — it runs on your Mac. Silence-only cuts make no AI call.
• Longer recordings and extra passes cost more, and Second Look adds another call.

Published rates change, so check your provider's current pricing page rather than trusting any figure quoted elsewhere.

Where keys are stored

Keys go into the macOS Keychain, encrypted by the system — never into a settings file and never into an exported configuration. RoughCut sends a key only to the provider it belongs to: OpenAI, Google or Anthropic. There is no RoughCut server in the path.

In Settings → Basic (or Full → LLM) each key field shows Not set with a paste prompt, or a green confirmation with the value masked once saved, alongside Replace… and Remove. The stored value is never displayed back to you.

Check a key before a long job

• Choose Model & Estimate… → Refresh Models asks the provider for its model list with your saved key. If the list fills, the key is valid and reachable; if it fails, the key is wrong, expired, revoked, or unfunded.
• If a job fails at the AI stage, the error names the provider and the cause: no key saved, a quota or billing problem, an invalid key, or a timeout. The provider-specific tables in Create and add an OpenAI API key, Create and add a Gemini API key and Create and add a Claude API key map each message to its fix.
• Run a short test clip with Bad-take removal on before committing an hour-long recording.

Common mistakes

• Assuming a provider is free. The default Gemini model is a paid-tier model, OpenAI API usage is billed separately from any ChatGPT plan, and Claude API usage is billed by Anthropic.
• Assuming Gemini is still the default. It was in earlier versions; new installations now default to OpenAI. Your saved choice survives updates either way.
• Clicking Use Model and expecting the next job to use it. Use Model only fills the field. Save Configuration applies it.
• Raising the reasoning level to get a better cut. A higher level takes longer and costs more, and it does not always cut better.
• Confusing the transcription source with the editing provider. They are separate settings. Choosing Gemini or Claude for editing does not change how audio is transcribed.
• Adding a key you do not need. For silence-only cuts, skip it.
• Keeping copies of your key in notes or chats. Treat it as a password; rotate it at the provider if it leaks.

Next steps

• Create and add an OpenAI API key
• Create and add a Gemini API key
• Create and add a Claude API key
• Bad-take removal and aggressiveness