dryvist-openrouter-models

name	dryvist-openrouter-models
description	Discover current leading OpenRouter models and live prices (public keyless API), select an escalation model for complicated work or advanced coding, call it through the LiteLLM router within a $1/day budget, and request onboarding of models the router does not serve yet
version	1.0.0
author	dryvist homelab
license	MIT
platforms	["linux"]
metadata	{"hermes":{"category":"research","tags":["openrouter","models","pricing","escalation","budget","dryvist"],"related_skills":["dryvist/github-issues"]}}

dryvist openrouter-models

Your resident local brain handles routine work. For complicated reasoning or advanced coding where a stronger frontier model would genuinely change the outcome, you may escalate to an OpenRouter model — a deliberate per-call choice, never an automatic on-error fallback. This skill teaches you to find what is currently available, what it costs, and how to use it within budget.

Budget: $1.00 per day (hard rule)

• Track your spend in memory under key openrouter-spend-<YYYY-MM-DD>: after each paid call, estimate cost from the response's token usage times the per-token prices you discovered, and add it to the day's total.
• Before any paid call, check the day's total. At or above $1.00: use the resident brain or a :free variant instead, and note the deferral.
• Prefer :free variants whenever they are adequate for the task. Never run a long unattended loop on a paid model.

Discover current models and prices (keyless, public)

The OpenRouter model catalog is public — no API key involved:

curl -s https://openrouter.ai/api/v1/models | jq '
  .data[] | {id, context_length,
             prompt_usd_per_mtok: ((.pricing.prompt | tonumber) * 1000000),
             completion_usd_per_mtok: ((.pricing.completion | tonumber) * 1000000)}'

Useful selections:

• Cheapest strong coders: filter ids matching your task domain, sort by completion_usd_per_mtok, read the top few.
• Free tier: .data[] | select(.id | endswith(":free")) | .id.
• For "what is currently leading", combine this catalog with a short web check (your web toolset) of OpenRouter's public rankings — the catalog is ground truth for price and context length; rankings are opinion.

Refresh discovery when you actually need it (prices change), not on a schedule. Cache the day's findings in memory rather than re-fetching per call.

Use a model — always through the router

Every model call goes through the homelab LLM router (your normal endpoint). The OpenRouter credential lives in the router, never on this guest. Request the model by its exact id as the router serves it. Registered today:

• deepseek/deepseek-v4-flash — paid, 1M context. The workhorse escalation.
• nvidia/nemotron-3-ultra-550b-a55b:free — free tier, rate-limited, and the provider logs prompt/session data on :free — never send confidential material through it.

A model id the router does not serve will fail — that is the signal to use the request lane below, not to retry blindly.

Request a model the router does not serve yet

Onboarding a new OpenRouter model is an infrastructure change (a per-model credential plus a router registration), so it is not self-serve. When discovery shows a model clearly worth having, file ONE GitHub issue in dryvist/ansible-proxmox-ai titled [hermes-model-request] <exact-model-id> containing: the exact OpenRouter id, current prompt/completion price per 1M tokens, context length, the concrete task class that justifies it, and the expected spend inside your $1/day budget. Check for an existing open request first — never duplicate. Real model ids only; never propose a generic alias.

Attribution

State the model(s) you actually used in every delivered message (see your persona's attribution rule) — the resident brain by name when you did not escalate, and every escalation model when you did.