Task Routing
Different AI tasks have different requirements. Photo analysis needs a multimodal model, while text extraction and simple commands work fine with a smaller, cheaper one. Task Routing lets you set the provider, model, and endpoint URL per task group.
OpenBin groups AI tasks into three categories:
| Task Group | Tasks Included | Typical Model Choice |
|---|---|---|
| Vision | Photo analysis | A multimodal model (e.g. gpt-5-mini, gemini-3-flash-preview) |
| Quick Text | Commands, execute, text extraction/structuring | A fast, cheap model (e.g. gpt-4.1-mini, claude-haiku-4-5) |
| Deep Text | Inventory queries, reorganization | A capable reasoning model |
Each group can override any combination of provider, model, and endpoint URL. Unset fields inherit from your default AI configuration.
Example configurations
OpenAI
| Group | Model |
|---|---|
| Vision | gpt-5-mini |
| Quick Text | gpt-4.1-mini |
| Deep Text | gpt-5-mini |
Anthropic
| Group | Model |
|---|---|
| Vision | claude-sonnet-4-6 |
| Quick Text | claude-haiku-4-5 |
| Deep Text | claude-sonnet-4-6 |
Google Gemini
| Group | Model |
|---|---|
| Vision | gemini-3-flash-preview |
| Quick Text | gemini-2.5-flash |
| Deep Text | gemini-3-flash-preview |
Mixed-provider example
You can route different task groups to entirely different providers. For example, use Gemini for photo analysis and Anthropic for everything else:
| Group | Provider | Model |
|---|---|---|
| Vision | Google Gemini | gemini-3-flash-preview |
| Quick Text | Anthropic | claude-haiku-4-5 |
| Deep Text | Anthropic | claude-sonnet-4-6 |
How to configure
- Go to Settings → AI → Task Routing.
- Select a task group (Vision, Quick Text, or Deep Text).
- Optionally override the provider, model, and endpoint URL for that group.
- Save.
Leave all fields blank for a group to use your default AI configuration.
TIP
One model for all tasks is a fine starting point. You can always add per-group overrides later — e.g. a cheaper model for Quick Text.
Server-wide task routing
Admins can lock task routing for specific groups via environment variables. When a group is configured by the server, it appears as read-only in the UI and cannot be changed by individual users. See the Configuration Reference for the full list of per-group variables.
Env-locked AI keys
When the server-wide AI_API_KEY (or one of the per-group AI_VISION_API_KEY / AI_QUICK_TEXT_API_KEY / AI_DEEP_TEXT_API_KEY keys) is set, the corresponding fields appear masked and read-only in the user UI; users cannot edit or delete them. This is how self-hosted admins provide AI to all users without storing individual keys.
Cloud AI credits
On the cloud product, each plan includes an AI credit budget on a rolling 30-day window — the period resets 30 days after your first request, not on the calendar month. Self-hosted instances are not credit-limited; AI calls go directly to your configured provider.
Cost is flat per unit — no formulas, no batch discounts:
| Class | What it covers | Credit cost |
|---|---|---|
| Quick text | Ask, command, query, structure-text, correct | 1 per request |
| Vision | Photo analysis, reanalysis | 5 per image — so 1 photo = 5, 2 = 10, 3 = 15 |
| Reorganize | Reorganize, tag suggestions | 2 per bin — so 7 bins = 14, 30 = 60, 100 = 200 |
Headline budgets:
- Free: 30 credits/month
- Plus: 100 credits/month
- Pro: 700 credits/month
- Trial: 30 lifetime credits (no monthly reset). CSV export is available during the trial; full JSON/ZIP export unlocks once you subscribe.
Credits are debited atomically before each request and refunded automatically if the upstream provider fails before producing a usable result. Your remaining credits and reset date are visible in Settings → AI, and the in-app composer surfaces the per-request cost before you submit.
Network security (self-hosted)
AI provider calls run through an allowlist with DNS pinning to prevent SSRF attacks against internal network resources. Self-hosted mode relaxes this to allow private addresses, so local endpoints like Ollama (http://localhost:11434/v1) work without configuration.