Compute Ladder

This is a cost-first model routing strategy that keeps you on local Ollama models (qwen3-coder, gpt-oss) unless infrastructure actually fails. It defines five tiers from free local to paid Claude Opus, but the philosophy is explicit: you only climb the ladder on 429s or timeouts, never because a task looks hard or you want better output. Tier 1 uses Cerebras at up to 2100 TPS for free, tier 3 brings in Hermes 405B for deep reasoning, and tier 4 is break-glass only with spending alerts at $5 daily. If you burn through API budgets or want to stay on local models by default with automatic failover only when services are down, this is the routing logic. It includes health checks and a spend guard that flags you when OpenRouter usage crosses thresholds.

GitHub stars: 2.8K
last updated: May 28, 2026

View on GitHub

Install

$ npx -y skills add open-gitagent/opengap --skill compute-ladder --agent claude-code

Installs into .claude/skills of the current project.

Files

Compute Ladder

Tier Definitions

Tier 0 — Local (never dies, zero cost)
  ollama/qwen3-coder:latest    primary, MoE, 128 TPS, 64k ctx
  ollama/gpt-oss:20b           fallback, 32k ctx HARD LIMIT

Tier 1 — Fast Free Cloud (up to 2100 TPS)
  cerebras/qwen-3-235b-a22b-instruct-2507   235B MoE, fast free
  cerebras/llama3.1-8b                      8B, ultra-fast light tasks

Tier 2 — Free Cloud (normal latency)
  openrouter/z-ai/glm-4.5-air
  openrouter/qwen/qwen3-coder

Tier 3 — Free Cloud Deep Reasoning
  openrouter/nousresearch/hermes-3-llama-3.1-405b:free

Tier 4 — Break-Glass (paid, restricted use)
  openrouter/anthropic/claude-opus-4.6  [narco-check and audit ONLY]