Meeting summaries through OpenRouter: one key, any model

Autorec records and transcribes your calls on your own machine. The summary is the one optional step that talks to a language model, and you choose which one. If you don’t want to open an account with OpenAI, then Anthropic, then Google just to compare them, OpenRouter gives you one API key and one bill for all of them.

What OpenRouter is, in one paragraph

OpenRouter is an API gateway. You send an OpenAI-style request to https://openrouter.ai/api/v1, name a model like openai/gpt-4o-mini or anthropic/claude-sonnet-5.5, and it routes the call to a provider that serves it. Autorec speaks that same chat-completions format, so OpenRouter works as a drop-in endpoint. Per OpenRouter’s FAQ, inference itself carries no markup; the fee is on buying credits (5.5% by card, 5% by crypto, as of this writing).

The trade-off: your transcript text leaves your machine and goes through OpenRouter to whichever provider runs the model. Audio and video never do. If that matters for your calls, read the privacy side of OpenRouter before you pick a model, or use a local model instead.

Set it up in Autorec

1. Create an OpenRouter account and add credit. Top up on the credits page in your account settings. A few dollars lasts a long time at the prices below.
2. Create an API key. Use the keys section of your OpenRouter account settings and copy the key somewhere safe. You only see it once.
3. Open Settings in Autorec and scroll to AI Summary. Under Provider, pick OpenRouter from the Aggregators group. The preset fills in the API base URL (https://openrouter.ai/api/v1) and a default model.
4. Paste the key into API Key, and set Model to the model you want. Model names use OpenRouter’s vendor/model form, for example anthropic/claude-haiku-4.5.
5. Click Save Settings. Then open any transcribed recording and press Generate Summary, or switch on Auto-summarize after transcription so it happens on its own.

The summary lands next to the video in your library. The default prompt asks for a title, a one-line gist, a Decisions section and an Action items section with owners where they were stated. You can rewrite it; the action items prompt post shows how.

What Autorec actually sends

Autorec sends a system prompt and your transcript as two chat messages, and asks for at most the number you set in Max summary length (tokens) (1000 by default; the field accepts 100 to 2000).

Autorec sends the whole transcript. Only a very long call (past roughly 400,000 characters, about eight hours of speech) gets cut, with a ”[Transcript truncated]” marker at the end.

Which models do well on meeting transcripts

A one-hour transcript is around 11,500 tokens, so every model below has room to spare (all have at least a 128k context window). What differs is format-following and how a model copes with messy speech.

• Cheap and good enough: openai/gpt-4o-mini (the preset default) and deepseek/deepseek-v4.1-flash. Both produce a clean gist and bullet lists for a standup or a client call. Expect occasional misattributed owners.
• Mid-range: anthropic/claude-sonnet-5.5. Better at separating a real decision from a suggestion someone floated, and at keeping owners straight when several people talk over each other.
• Frontier: anthropic/claude-opus-5.5. Worth trying for a high-stakes call you will act on. For routine meetings the gain over a mid-range model is small.

This is my read from how these tiers usually behave on this task, not a benchmark. Run one real transcript through two models and compare. On the cheap models that costs a fraction of a cent.

One practical tip: reasoning models can spend part of the output cap on thinking tokens. If a summary comes back empty or cut off, raise Max summary length (tokens) or choose a non-reasoning model.

What a summary costs

Assumptions: a one-hour meeting at 130 to 150 spoken words a minute is roughly 8,000 to 9,000 words, or about 11,500 tokens. Add about 110 tokens of default prompt and you get 11,600 tokens in. I assume a 500-token summary back. Prices are from OpenRouter’s live model list, per million tokens, as of October 2026.

Model	Input / output per 1M tokens	Cost per summary (11,600 in, 500 out)
deepseek/deepseek-v4.1-flash	$0.0472 / $0.12	about $0.0006
openai/gpt-4o-mini	$0.15 / $0.60	about $0.002
anthropic/claude-sonnet-5.5	$2 / $10	about $0.028
anthropic/claude-opus-5.5	$4 / $20	about $0.056

One row worked out: Sonnet is 11,600 × $2 / 1,000,000 = $0.0232 for input, plus 500 × $10 / 1,000,000 = $0.0050 for output, which comes to $0.0282.

A hundred one-hour meetings cost about six cents on DeepSeek, twenty cents on GPT-4o mini, $2.82 on Sonnet and $5.64 on Opus. Input dominates, so a longer call costs proportionally more and the output stays the same.

Free models

OpenRouter lists models with a :free suffix, such as google/gemma-4-31b-it:free. According to OpenRouter’s FAQ, free usage is limited to 50 requests a day without credits, or 1,000 a day once you have bought at least $10 of credit. That is plenty for a handful of meetings a day.

On privacy, I could not find a statement in OpenRouter’s documentation about how free endpoints handle prompts. What it does say is that accounts have separate privacy settings for paid and free models, and that you can control whether requests route to providers that may train on your data. Until you have checked those settings yourself, assume a free endpoint may keep or train on what you send, and don’t put a client call through one. The OpenRouter privacy post walks through the settings.

Autorec also has an Extra request fields (JSON) box under AI Summary. With the OpenRouter preset selected, the Use privacy settings button fills it with {"provider": {"zdr": true, "data_collection": "deny"}}, which asks OpenRouter to route only to zero-data-retention providers that don’t collect your data. The privacy post explains what that does and what it limits.

A sensible starting point

Start with the preset’s default, openai/gpt-4o-mini, on one of your own standups. If the owners or decisions come out wrong, move to anthropic/claude-sonnet-5.5. Keep Opus for the calls where a missed action item has a price tag. Switching is one field in Settings, and your key stays the same.

Autorec is a one-time purchase, so the summary is the only running cost. See features for what else the app does, and the configuration docs for the settings.