← Back to feed

OpenRouter’s April release dropped four infrastructure upgrades in parallel, reflecting a deliberate push to absorb every AI modality under one routing and billing layer. The changes compound on each other: video generation, an agentic SDK, multi-project isolation, and reranking now all operate under the same API keys and observability stack a team already uses for text.

Video Generation

OpenRouter now supports text-to-video and image-to-video generation through a unified API. Day-one models include Seedance 2.0, Veo 3.1, Wan 2.7, and Sora 2 Pro. The implementation adds normalized parameters, async job tracking, and capability discovery — the same patterns OpenRouter uses for text models, applied to video.

Video sits alongside text, images, audio, embeddings, and rerankers under a single routing and billing layer. The practical effect is that a team can move between modalities without managing separate vendor accounts, rate limit budgets, or billing systems.

Agent SDK

@openrouter/agent ships callModel: a single function handling multi-turn tool calling, streaming, stop conditions, and cost tracking across 300+ models. Tools are defined with Zod schemas. Stop conditions use readable predicates: stepCountIs(10) or maxCost(1.00).

Two scaffolds ship on top: create-agent-tui generates a coding agent with a terminal UI; create-headless-agent builds a multi-model CLI tool. The SDK is notable because it exposes cost tracking as a first-class stop condition — not a bolt-on — which addresses one of the more common failure modes in production agent deployments.

Workspaces

OpenRouter Workspaces let teams partition projects into isolated environments, each with its own API keys, routing defaults, guardrails, and observability. The feature targets multi-project developers and enterprise teams needing staging and production isolation. Existing accounts are migrated into a Default workspace with no breaking change.

Reranker Models

A new /rerank endpoint supports Cohere Rerank and Fireworks rerankers, routing retrieval and document-chunk reordering through the same infrastructure. Combined with Model Fusion — which aggregates responses from multiple models into a single output — the April release pushes OpenRouter further into the retrieval pipeline rather than purely at the generation layer.

Additional Shipping

• Prompt history: View, replay, and remix previous prompts; replay the same input across different models for output comparison
• Benchmark comparisons on model pages: Design Arena ELO rankings now visible on any model’s comparison page
• Knowledge cutoff dates in /models API: Programmatic access to training cutoff dates across all models
• Stripe Projects: stripe projects add openrouter/api from CLI creates an OpenRouter account, API key, and billing — an agent-executable setup flow
• Typed tool context (TypeScript SDK): contextSchema on tools with setContext() mutation, Zod-validated, persisting across turns

Frontier Models in April

OpenRouter’s April model onboarding included GPT-5.5, DeepSeek V4 Pro and V4 Flash, Claude Opus 4.7, GPT-5.4 Image 2, and Kimi K2.6 — the majority of the major frontier drops from the month, now accessible under a single API key.

The strategic picture: OpenRouter’s value proposition has shifted from “single endpoint for 300 text models” to “single endpoint for every AI modality, with built-in agent orchestration.” The video and reranker additions are the clearest indicators that the company is positioning itself as infrastructure for compound AI systems, not just model selection.