Feedoptimise now connects to OpenRouter, adding the widest model choice yet to our multi-model AI product feed optimization. One API key gives your feed access to more than 400 models from over 60 providers, including a rotating set that costs nothing per token. Pick one of those, point it at your titles, descriptions or categories, and you can optimize your catalog without paying for inference.
What is OpenRouter?
OpenRouter is a routing layer that sits in front of most of the commercial and open-weight language models on the market. Instead of holding separate accounts and separate keys with OpenAI, Anthropic, Google, DeepSeek, xAI, Meta, Mistral, Qwen and everyone else, you hold one OpenRouter account and address every model through a single API.
The catalog currently runs to more than 400 models across 60-plus providers. Some are the same flagship models you would get direct from the provider. Some are open-weight models hosted by third parties. And roughly two dozen at any one time are published as free variants, priced at $0 per million input and output tokens.
Why this matters for feed optimization
Feed work covers a dozen jobs of differing difficulty, and they do not all need the same model.
Extracting a color from a description is a simple classification task. A small, cheap or free model handles it correctly almost every time. Rewriting 80,000 titles to a channel-specific pattern while respecting character limits and brand rules is harder. Translating product copy into Nordic languages while keeping the product terminology intact is harder again, and quality varies between models.
Until now, matching each job to the right model meant signing up with each provider, generating a key, and adding it separately. With OpenRouter you add one API key, then choose the model per modifier. Title rewrites can run on a frontier model, attribute extraction can run on something cheap or free, and you can change your mind next week without touching your integration setup.
The other practical benefit is testing. Benchmarks will not tell you which model writes better titles for your catalog. You find that out by running it on a sample of your own products and comparing the output. With every model one dropdown away, that takes minutes.
Free models, and what free actually means
Free models are free on tokens and capped on volume. OpenRouter allows 50 free calls a day on an account that has never bought credits, and a one-time top-up of $10 or more lifts that to 1,000 a day for good. Each item is optimized on its own, so a thousand calls means roughly a thousand products a day per prompt. There is a second limit of 20 calls a minute, which paces a long run rather than stopping it.
None of that draws down your balance. Free variants stay at $0 per token, so the credit you add sits there for your paid models while the free work stays free.
A free model costs you time rather than money. That rules out rewriting 80,000 titles in an afternoon, and leaves most other feed jobs open.
Jobs free models handle well:
- Prompt development, so you pay nothing while you iterate on the wording and see the output on real products
- Model comparison, running the same prompt through several models on a sample of your catalog before committing to one
- Title and description rewrites on catalogs of a few thousand items, or on larger ones where a run spread across several days is acceptable
- Targeted work on a filtered subset, for example only the items your Analysers flagged as having weak titles
- Ongoing enrichment of new arrivals rather than a full re-run of the catalog
- Translations for a new market, where the alternative is either a human translator or a paid model
Free models pair well with the way Feedoptimise caches AI output. We hold generated content and refresh it when your source data or prompt settings change, so you pay to optimize a title once rather than on every feed refresh. Put a zero-cost model behind that and a catalog can stay optimized without an inference bill at all.
On quality, the free roster is open-weight only, and recent releases from labs like DeepSeek, Qwen, Meta, Mistral and Google are strong enough to be worth testing for title rewriting, category mapping and translation. Whether one of them beats a paid model on your products is a question for your own sample, not for a benchmark table.
Privacy: the free tier has a condition attached
Most providers serve free endpoints on the condition that they can train on the prompts they receive, and some publish prompts and completions to public datasets. OpenRouter controls this with two settings under Settings, Privacy: free endpoints that may train on request data, and free endpoints that may publish prompts. If either is switched off, free models return a 404 saying no endpoints match your data policy, which is the most common reason a free model looks broken when it is not.
That decision sits with you. Our job is to give you the connection and let you choose the model. If your product data is not sensitive, free models are an easy win. If it is, use a paid model on OpenRouter, use one of your existing direct integrations, or use a self-hosted open-source model where the data never leaves the infrastructure. This is a difference of degree rather than kind. The major paid APIs do not train on your data by default, and most of them still let you opt in if you want to.
How the integration works
The setup follows the same pattern as our other AI integrations:
- Create an API key in your OpenRouter account.
- Add the key to your Feedoptimise account under AI integrations.
- Create a new AI modifier in your feed and select OpenRouter as the provider, then pick the model you want from the catalog.
- Choose a predefined prompt or build a custom one, map it to a target attribute, and run it.
One detail specific to OpenRouter is worth knowing before your first large run. Many models are served by more than one upstream host, and OpenRouter falls back to an alternative automatically when one is unavailable or at capacity. On a long optimization run across a large catalog, that cuts down the number of partial failures you have to go back and re-run.
Because you connect with your own key, you pay OpenRouter directly at their published rates. There is no markup from Feedoptimise and no separate AI credit system to manage. Usage, spend and remaining quota stay visible in your own OpenRouter dashboard.
Predefined prompts, ready out of the box
If you want results without writing a prompt, the predefined library covers the most common feed optimization jobs. The Title Optimizer restructures product titles around the attributes that matter for each channel, such as brand, product type, color and size, so your items match more of the search queries buyers actually type. Point it at OpenRouter, choose a model, and run it. Predefined prompts also work as a starting point: load one, adjust the wording for your catalog, and save it as your own.
Custom prompts, full control
For anything the predefined prompts do not cover, the visual prompt template editor lets you write your own instructions and insert any feed attribute as a variable. Common builds include:
- Title rewrites enforcing your own pattern per channel, for example Brand + Product Type + Color + Size, with high-performing keywords injected
- Product type optimization that pushes relevant keywords into the product_type attribute, which carries real ranking weight on Google Shopping
- Description rewrites that follow brand tone-of-voice guidelines
- Category mapping against a channel taxonomy or a custom keyword-rich structure
- Attribute extraction, pulling color, material or gender out of unstructured description text
- Feed translations for entering new markets
- Data quality audits, for example asking the model to rate title relevance from 1 to 10 and flag items below a threshold
You write the prompt once, and Feedoptimise applies it to selected items or the whole catalog. Because the model is a per-modifier setting, you can run the same prompt through two different models and compare the output side by side.
Verifying that the output performs
A better title on paper still has to earn the click and the conversion. Push OpenRouter-generated titles or descriptions into the product feed A/B Testing Suite and let the platform apply the winning variant automatically. If a free model produces titles that convert as well as a frontier model on your catalog, the test will show it, and you can keep the cheaper option with evidence behind the decision.
Getting started
OpenRouter support is available now on all plans. Log in, open your AI integrations, add your OpenRouter API key, pick a model, and run your first prompt. If you already run DeepSeek, OpenAI, Claude or Gemini through Feedoptimise, those integrations continue to work as before. OpenRouter is an additional option, useful when you want breadth of model choice from a single API key.
Not using Feedoptimise yet? Start a free trial or book a live demo and we will walk you through it.
