OpenRouter

A practical walkthrough for scoring agent outputs with a second model against a written rubric, covering judge-vs-candidate bias pitfalls, scoring modes, and how to keep judging cost and reliability in check.

Step-by-step configuration for pointing Cursor at OpenRouter through the OpenAI key and base URL override, using Moonshot's Kimi K2 as the worked example, including which Cursor warning is safe to ignore.

OpenRouter's postmortem on two February outages: an external caching layer dropped database connections, mass invalidation overwhelmed lookups, and timeouts were returned as 401 auth errors. A parallel DoS attack led the team to misdiagnose the first incident.

OpenRouter added video generation to its unified API: text-to-video and image-to-video across Seedance, Veo 3.1, Wan and Sora 2 Pro, with async job tracking, parameters normalized across models, and a models endpoint that reports what each one supports before you call it.

OpenRouter replaced its fixed Stripe charge with percentage platform fees: 5.5% and an $0.80 minimum on card payments, 5% flat on crypto, with a published comparison of old and new fees across order sizes.

OpenRouter launches a benchmark page scoring model routers on quality, speed and cost across six benchmarks, with a blended Router Index. It also explains why routing often loses to a single model: cache rebuilds, weak complexity signals and added latency.

Workspaces split one OpenRouter account into isolated environments, each with its own API keys, routing defaults, guardrails, presets, plugins and observability. Billing, logs and privacy restrictions stay account-level, and existing setups land in a Default workspace.

MiniMax's open-weights Hailuo H3 is on OpenRouter: a lightweight video model taking text, image, audio and video input with native audio-visual output, built for instruction-led edits, text and brand rendering, and video-to-video motion transfer. Up to 12 reference files per request.

A TypeScript walkthrough for building a tool-calling agent loop: executing tool calls by ID, capping iterations, handling repeated failed calls, and adding model fallback.

AssemblyAI's Universal-3.6 Pro speech-to-text model is on OpenRouter. It claims 30% fewer errors on short conversational utterances, 12% fewer overall, better formatting of codes and addresses, and 14 new languages.

A method for judging serving environments rather than models: latency, throughput, uptime and quantization, read at the tail instead of the average. The core point is that a shared model slug hides precision and infrastructure differences that show up in production quality.

OpenRouter and a16z publish an empirical study of a year of gateway traffic, more than 100 trillion tokens, tracking the move from single-pass generation to reasoning models and agentic workflows and the growing production use of open-weight models.

OpenRouter data shows a new model launch mainly cannibalized flash-tier models across labs, and nearly half of its users on the platform hadn't touched any model the week before its launch.

OpenRouter opens its first image endpoint with Gemini 2.5 Flash Image Preview: $0.30 per million input tokens, 32k context, multiple images per call, and character consistency across scenes for work like comic panels and bulk logo iterations.

A CLI that boots Claude Code, Codex, OpenCode, or Hermes against OpenRouter with the right configuration per model. The notable disclosure is the env matrix itself, particularly tool search, which Claude Code enables only for the first-party harness and which materially changes token use and turn count.

Pointing OpenAI's Codex CLI at a third-party gateway takes one config.toml block plus two constraints that break it silently: the Responses wire protocol is now mandatory, and a project-local config is ignored for provider settings. Includes model pricing and the model_not_found fix.

OpenRouter lists Gemini 4 Argon via Google AI Studio and Google Cloud at introductory $2/M input and $10/M output, reporting first place on AutomationBench and LVBench and a tie on CWE-bench v1.

BYOK is generally available on OpenRouter: attach your own provider keys to combine rate limits, spend existing cloud credits, and keep all model usage in one dashboard, billed at 5% of the upstream provider's cost.

OpenRouter launches US In-Region Routing alongside its EU routing: requests to region-specific endpoints are decrypted and served only by in-region providers, enforceable per-workspace via Guardrails.

OpenRouter's changelog for Ori covers the TUI, Harness, and Eval in one place. Release 0.10.1 fixes variant models being treated as unknown, clamps misleading zero prices in the model picker, drops batch-only endpoints from interactive selection, and allows a custom palette file.

OpenRouter breaks down what Zero Data Retention actually covers: provider-side storage of prompts and responses, not data in transit, your own logging, or third-party tools. Explains how to enforce ZDR routing via account settings or the provider.zdr request field.

OpenRouter walks through committing a fixed eval set, scoring it with a script, measuring noise from repeated runs, and wiring it into GitHub Actions as a required check so prompt regressions cannot merge.

Adding a gateway as a provider in the Kilo Code editor extension, and the config file that controls it: the exact nesting path for provider routing rules, sorting by price, throughput or latency, data-policy fields, and the middle-out transform for prompts past the context window.

OpenAI cut prices on GPT-5.6 Terra and Luna, and OpenRouter's 50% exclusive discount stacks on top: Luna lands at $0.1/M input and $0.6/M output, Terra at $1/M input and $6/M output, pushing Luna further onto the cost-quality frontier.