OpenRouter

The argument for treating agent governance as a runtime control rather than a policy document: budget caps, approved-model restrictions and audit logging sit on the API key, because that is the one chokepoint every autonomous request must cross. Includes the common escalation failure modes.

Ori Eval runs inside your coding agent, finds where the codebase calls models, asks what you optimize for, then writes an eval and ranks candidate models against it. Assertions cover which tools were called, which were avoided, and whether the run completed, with LLM-as-judge for open-ended answers.

How to turn production traffic into a reviewed, Git-versioned regression set for LLM behavior. Covers sizing (about 10 items to explore, 100 to 1,000 for regression), failure-mode coverage, CI runs, and benchmarking candidate models on your own data.

A comparison of server-side code execution tools from OpenAI, Anthropic, Google and OpenRouter, covering sandbox isolation, persistence between requests, network defaults, per-second pricing and cases for self-hosted sandboxes.

A benchmark of image input detail levels across five vision models finds the cost-saving move backfires on reasoning models: accuracy drops sharply and reasoning tokens climb enough to erase the input savings. Non-reasoning models still benefit, and reasoning effort is the better lever.

OpenRouter's tutorial shows how to edit images with Gemini's Nano Banana 2 through one API call: send a source image plus prompt, decode the result, and iterate by feeding outputs back as new inputs.

How to point Claude Code at OpenRouter with three environment variables and no proxy: the Anthropic-compatible endpoint passes thinking blocks, native tool use and streaming through untouched, while /status confirms the switch and one key fronts shared billing and per-developer spend caps.

A tutorial on cheap-first model routing for support bots. It compares static rules, classifier triage, and answer checks, clarifies which parts OpenRouter handles, and shows how to evaluate whether each escalation to a stronger model earns its cost.

A model-by-model breakdown of which DeepSeek V4 variants accept images: only V4.1 Flash and the experimental Vision Exp checkpoint do, while Pro 0813, Flash 0731, and older checkpoints stay text-only.

A working reference for Gemini 2.5 Flash: per-million pricing across AI Studio, Vertex and OpenRouter, how the thinking budget behaves at zero and in dynamic mode, document limits of 50MB and PDF or plain text only, and the absent capabilities such as image generation.

Running stripe projects add openrouter/api provisions an OpenRouter account, generates an API key into an encrypted project vault, syncs it to .env and attaches Stripe billing from the command line. Existing accounts can be linked, and 20+ other providers use the same flow.

OpenRouter streams cost and token accounting inline when you opt in with usage.include, breaking out cached prompt tokens and reasoning tokens per response. Activity charts also respect table filters, and SVG output previews in the chatroom.

Web search and page fetch as server-side tools any tool-calling model can invoke, with one schema across providers, a choice of native, Exa, or Parallel backends at $0.005 per request, domain allow and block lists, and caps on parallel searches per agent loop.

Two providers behind OpenRouter's free inference moved to paid-only. The gateway is onboarding replacements and covering some cost itself, keeping DeepSeek V3 and R1 free while dropping lower-traffic models from the free tier.

OpenRouter added a hosted Linux sandbox tool (openrouter:shell) to its Responses API, letting any model on the platform write and execute code with results read back, without users hosting their own execution environment.

An MCP server that hands your coding agent live model data, including catalog, third-party benchmark rankings, per-provider pricing and latency, docs search, and test inference, behind a dedicated capped key. Install commands are included for the major agent clients.

A tool type that decides per call whether an agent proceeds or waits for a person: return a value to auto-resolve, return null to pause the loop, then resume with the human decision as a function call output. Includes a hook for validating that answer first.

Muse Spark 1.3 is now served on OpenRouter, positioned for long-running agentic, multi-agent, and coding workflows, with a lower cost Contributor tier live alongside it. The model is described as tracking what it learns and asking for confirmation on conflicting inputs.

OpenRouter turns on JSON Schema validated structured outputs for OpenAI 4o and Fireworks models, with more providers promised, and opens free but tightly rate-limited access to Gemini Flash 2.0 and its faster time to first token.

OpenRouter now repairs malformed JSON from structured-output calls before it reaches your app, fixing trailing commas, unescaped control characters and missing brackets. Published per-model valid-rate tables show defect reductions from roughly 10% to 99.8%.

Usage analytics for teams running agents: group spend, tokens, cache hit rate and P50/P90/P99 latency by model, key, app, session or your own classifier dimensions, then click through to the generations behind any bar. Same data exposed via an Analytics API.

The openrouter:advisor tool lets a cheap executor model consult a stronger model mid-generation and keep going with the result. Either role can be any model, consultations are single-turn and depth-capped against recursion, and advisors reconstruct prior consultations on replay.

Thirteen routes to free model inference compared on request limits, context windows, and data-use policy, with three categories kept distinct: indefinite free tiers, expiring trial credits, and running open weights locally. Also covers what breaks when a free tier runs out.

Nous Research's Hermes Agent is an MIT-licensed terminal agent that runs on any backing model. The guide covers its gateway setup, the ~64K context minimum imposed by system prompt plus tool schemas, provider routing and fallback chains, and offloading auxiliary tasks to cheaper models.