OpenRouter Free Model Rotate v2.0#

Capabilities#

• Smart scoring — ranks models by context window, multimodal support, reasoning ability, output length, and latency
• Concurrent testing — tests multiple models at once (configurable workers)
• Quality benchmark — PONG instruction-following test
• Capability filter — text-only / multimodal / image / reasoning / large context
• Result caching — 1-hour cache to avoid redundant API calls
• JSON report — export results for analysis
• Auto config — updates openclaw.json + models.json + optional gateway restart

Quick Start#

bash

# Full flow: scan → bench → test → update → restart (recommended)
scripts/rotate_free_models.py --api-key "sk-or-xxx" --restart

# Quick rotate (no bench, just connectivity)
scripts/rotate_free_models.py --api-key "sk-or-xxx" --test 30 --keep 10 --restart

# Scan + show ranked by score (no changes)
scripts/rotate_free_models.py --api-key "sk-or-xxx" --scan --sort score

# Quality benchmark
scripts/rotate_free_models.py --api-key "sk-or-xxx" --bench --json report.json

# Filter: only multimodal models
scripts/rotate_free_models.py --api-key "sk-or-xxx" --filter multimodal --restart

# Use cached results (<1h old), skip retesting
scripts/rotate_free_models.py --api-key "sk-or-xxx" --use-cache --keep 10

# Save JSON report
scripts/rotate_free_models.py --api-key "sk-or-xxx" --json /tmp/report.json

Parameters#

FlagDefaultDescription--api-key``$OPENROUTER_API_KEYOpenRouter API key--test N``0 (all)Max models to test--keep N``10Working models to keep in config--workers N``5Concurrent test workers--timeout N``15Per-model timeout (seconds)--benchoffEnable quality benchmark--filter TYPE``allall/text/multimodal/image/reasoning/fast/large--sort BY``scorescore/latency/name--use-cacheoffUse 1h cached results--json FILEnoneSave JSON report--restartoffRestart gateway after update--no-updateoffDon't change configs--scanoffScan only, no testing

Scoring Algorithm#

Models are ranked by a weighted composite:

FactorWeightEffectContext window+2 per 100K tokensMore context = higher scoreMax output tokens+0.5 per 1KLonger output = higherImage input+5Multimodal bonusAudio input+5Multimodal bonusVideo input+3Advanced capabilityReasoning support+8Chain-of-thought bonusLatency-0.3 per 100msFaster = higher scoreBrand quality+2~5Qwen-Coder, Llama-70B, GPT, Gemini recognized

Scheduling#

Run via cron every 6 hours for auto-rotation:

text

0 */6 * * *  python3 rotate_free_models.py --api-key "sk-or-xxx" --restart > /var/log/model-rotate.log 2>&1