Auto Router

Create a product photograph of a minimalist ceramic coffee mug on a wooden surface with soft morning light. Clean composition, shallow depth of field, warm tones.

The router selects an image model optimized for photorealistic rendering and delivers a crisp product shot: the mug sits slightly off-center on weathered oak planks, its matte white glaze catching diffused sunlight from the left. The background blurs into warm bokeh. Shadows are soft and natural, with subtle color grading that emphasizes the cozy morning aesthetic. Surface textures—the wood grain, the ceramic's slight imperfections—read as authentic rather than over-processed.

Auto Router's strength here is model selection intelligence: it routes to whichever backend currently handles product photography best, so you get consistent quality without tracking individual model updates. The trade-off is opacity—you don't control which specific model runs, and pricing varies based on the router's choice, making cost prediction difficult for budget-sensitive workflows.

Generate an editorial illustration for an article about remote work burnout. Abstract, conceptual style with a person dissolving into digital particles. Muted blues and grays, slightly surreal.

The router identifies the conceptual-illustration use case and produces an image where a figure sits at a laptop, their upper body fragmenting into geometric shards and pixels that drift upward like smoke. The color palette stays restrained—slate blues, charcoal grays, one accent of burnt orange from the screen glow. The style balances abstraction with enough figurative clarity that the metaphor reads immediately. Composition follows editorial conventions: vertical orientation, negative space at top for headline placement.

This example shows the router handling stylistic requirements well—it interprets 'abstract, conceptual' and routes accordingly rather than defaulting to photorealism. However, the 2-million-token context window matters less for image generation than for multimodal analysis tasks, and without benchmark data, it's unclear how the router's style consistency compares to using a single known model repeatedly.

Design a fantasy book cover: a glowing sword embedded in ancient stone ruins at twilight, overgrown with luminescent vines. Dramatic lighting, painterly style reminiscent of classic fantasy art.

The router delivers a composition with strong genre awareness: the sword stands center-frame, its blade emitting cool white light that contrasts with the warm purple-orange sky. Crumbling stone archways frame the scene, their surfaces textured with moss and creeping vines that pulse with bioluminescent green. The painterly treatment shows visible brushwork and atmospheric perspective—distant ruins fade into mist. Lighting is theatrical, with the sword as the primary light source casting long shadows across fractured flagstones.

Auto Router's multimodal support (text, image, audio, file, video) means you could theoretically feed it reference images or audio mood boards alongside the text prompt, though this example uses text only. The main limitation remains pricing uncertainty—the '$?' placeholder reflects that costs depend on which backend model the router selects, making it harder to budget for high-volume cover design work compared to fixed-price alternatives.

When you need one endpoint for text, image, and video flags

A 4-person community platform runs 800 user uploads daily across photos, short videos, and text posts. Auto Router handles all three modalities through a single API call, routing each request to the best-fit model without manual switching. The 2M token context window means you can batch entire video transcripts with image frames in one pass. Pricing is opaque—OpenRouter doesn't publish per-token rates for the router—so you're trading cost visibility for routing convenience. If you're already juggling three separate moderation APIs and your volume justifies the integration overhead, this collapses your stack. Under 200 mixed requests per day, stick with dedicated endpoints where you control the model and the bill.

Fast iteration when you don't know which model you need yet

A 2-person startup is building a receipt-scanning expense tool and hasn't settled on OCR versus vision-language models. Auto Router lets them throw receipts, invoices, and handwritten notes at one endpoint while OpenRouter picks the backend model per request. The file and audio support means they can test voice memos and PDF uploads without rewriting the integration. No benchmarks exist for the router itself—you're inheriting whatever OpenRouter selects, which changes as they update routing logic. This works during the first 4-6 weeks of prototyping when speed matters more than cost control. Once you identify the 80% case, lock in a specific model to avoid surprise billing and latency variance.

When your team analyzes mixed media and needs maximum context

A 5-person policy research group processes 40-page PDFs with embedded charts, video testimony clips, and interview transcripts in a single analysis pass. Auto Router's 2M token context means they can load an entire hearing packet—documents, screenshots, audio transcripts—without chunking. The router picks the model, so they avoid the decision paralysis of choosing between GPT-4V, Claude, or Gemini for each media type. The trade-off: zero cost predictability and no benchmark transparency. If your workflow is already multimodal and you're spending 6+ hours per week on model selection, this buys back that time. If budget matters or you need sub-500ms latency, route manually to a known model.