OpenRouter

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.

Benchmarks TypeSafe's Jev decision model against GPT and Claude Opus on support triage, comparing accuracy, latency, and cost, and shows TypeScript patterns for routing with Jev, computing in code, and writing replies with an LLM.

Venice is now a provider on OpenRouter, leading with Dolphin Mistral 24B Venice Edition: a low-refusal fine-tune of Mistral 24B with no server-side storage, free to call through the API and chatroom.

Inception's Mercury Decide now accepts images in the state of a Decisions request on OpenRouter. It answers typed questions from the image with a choice, score, or yes/no plus probability. Pricing is unchanged and images bill as input tokens.

The arithmetic behind prompt caching in agent loops: what reads and writes actually cost per provider, why a single unreused write loses money, and why multi-provider routing quietly breaks the warm cache on turn two unless sessions are pinned.

OpenRouter's review of ByteDance's Seedance 2.5 breaks down its token-based per-second pricing (about $0.103/s at 480p, $0.231/s at 720p), 30-second clip length, image/video/audio reference inputs, and how it stacks up against Seedance 2.0, Wan 3.0, and Veo 3.1.

The OpenRouter Auto Router widens its candidate pool to 19 models chosen for quality, reports the model that actually served each request via API and UI, and pilots Not Diamond's meta model as the routing brain while access stays free.

Measured billing data from users who moved from GPT-5.4 to GPT-5.5 shows real cost rising 49 to 92 percent rather than the headline 2x, because reduced verbosity only offsets the increase on prompts above 10K tokens and reverses in the 2K to 10K range.

An argument about whether implicit prompt caching counts as retention under ZDR standards, tracing KV cache through GPU memory, host DRAM and NVMe paging, and explaining why Google classifies it as retention while most providers do not.

Cloudflare's Workers AI platform becomes an OpenRouter provider, adding edge-served open-weight options: Gemma 7B-IT with tool calling, DeepSeek R1 Distill Qwen 32B, and the Llama 3 through 3.3 line including an 11B vision model.

A method for model selection in agents: define the quality bar, run cheap, mid and frontier models on your own examples, compute cost per quality point from billed usage, and recheck as models and prices change.

OpenRouter's preset guide shows how to store a model list, system prompt, provider routing, and sampling parameters under one versioned slug (@preset/name) so every app referencing it updates when you edit the preset instead of redeploying code.

Routing splits into two independent decisions: which model answers, and which provider serves it. Every configuration knob maps to one of them. Covers the provider object, price-weighted load balancing, the throughput and cost shortcuts, fallback interaction, and where an auto router belongs.

Responses through OpenRouter now carry a normalized finish_reason following OpenAI conventions, with the provider's raw value preserved alongside as native_finish_reason. Code that switches on stop conditions works the same across models.

Fusion dispatches one prompt to a panel of models and has a judge model synthesize the responses. Across 100 DRACO deep-research tasks a Fable 5 plus GPT-5.5 panel scored 69.0 percent against 65.3 for the best solo model, and a budget panel beat both GPT-5.5 and Opus 4.8.

Models get pulled and restricted routinely, and a slug pinned in code turns that into a redeploy for every service. Covers moving model choice, fallback chains, provider rules and retention policy into a named server-side preset, and what rollback looks like once you do.

OpenRouter adds ElevenLabs Eleven v4, v4 Turbo and Scribe v2 among eleven voice models, callable with an OpenRouter key. Includes model selection guidance, character limits per request, and a 50% launch discount until October 19.

Quality-weighted provider routing becomes the default for tool-calling requests, rescoring providers every few minutes against throughput, tool-call schema validity measured across production traffic, and scheduled TauBench Verified and GPQA-Diamond runs.

Connecting the open-source OpenClaw agent runner to a model gateway: one onboarding command, the author/slug reference format with version pinning and routing suffixes, per-agent model assignment for cost control, and why server deployments must set the key in config rather than the shell.

OpenRouter documents how provider implementations of identical model weights diverge, shows benchmark spreads tightening as providers burn a model in, and introduces exacto endpoints that restrict routing to providers with measurably higher tool-calling success.

OpenRouter publishes a side-by-side image benchmark across 39 models, built on prompts chosen to break them: counting, negation, occlusion, exact typography, and minimal-diff edits. Results are visually judgeable and sortable by price and generation time.