Skip to main content
Astrolabe uses a static checked-in roster. Runtime routing does not depend on live discovery. OpenRouter is the maintainer reference for model IDs and pricing, but the active roster is whatever Astrolabe ships in src/manifests/index.js.

​
Core roster

KeyModel IDRole
m27minimax/minimax-m2.7Main workhorse for serious OpenClaw turns
m25minimax/minimax-m2.5Strict-budget, fallback, and overflow only
kimiK25moonshotai/kimi-k2.5Multimodal and visual specialist
kimiThinkingmoonshotai/kimi-k2-thinkingLong-horizon research/planning specialist
glm47Flashz-ai/glm-4.7-flashCheap first-line strict JSON and tool-argument model
glm5z-ai/glm-5Hard schema, tool, and system-design specialist
grokx-ai/grok-4.1-fastCheap long-context tool-capable model
dsCoderdeepseek/deepseek-v3.2Low-cost text-heavy overflow model
qwen35Flashqwen/qwen3.5-flash-02-23Cheap generalist for chat, retrieval, and summarization
qwen35Plusqwen/qwen3.5-plus-02-15Long-context multimodal research/value model
qwenCoderNextqwen/qwen3-coder-nextCost-sensitive coding specialist
gpt54openai/gpt-5.4Cross-family verifier or premium executor
gpt54Miniopenai/gpt-5.4-miniCheap control-plane and verifier role
gpt5Nanoopenai/gpt-5-nanoDefault router, classifier, and lightweight verifier
gpt54Nanoopenai/gpt-5.4-nanoRaw available only, not the default nano alias
gem25FlashLitegoogle/gemini-2.5-flash-liteStable cheap multimodal ingest/extraction model
gem25Progoogle/gemini-2.5-proStable premium multimodal branch
gem31FlashLitegoogle/gemini-3.1-flash-lite-previewPreview multimodal fast path
gem31Progoogle/gemini-3.1-pro-previewPreview multimodal research branch
grok420x-ai/grok-4.20Selective preview premium research branch
mimoV2Flashxiaomi/mimo-v2-flashRaw-only evaluation model
mistralSmall4mistralai/mistral-small-2603Raw-only evaluation model
sonnetanthropic/claude-sonnet-4.6Safe-lane default and premium escalation
opusanthropic/claude-opus-4.6Final escalation

​
Compatibility aliases

AliasResolves To
nanogpt5Nano
minigpt54Mini
gemFlashgem25FlashLite

​
Virtual models

These are the recommended public-facing models for OpenClaw users:
  • astrolabe/auto
  • astrolabe/coding
  • astrolabe/research
  • astrolabe/vision
  • astrolabe/strict-json
  • astrolabe/cheap
  • astrolabe/safe

​
Roster policy

  • m27 is the default workhorse.
  • m25 is intentionally not the main default anymore.
  • gpt-5-nano owns the control plane by default.
  • Kimi and GLM stay first-class specialists.
  • Preview and experimental models are opt-in and excluded from default lane manifests.
  • grok420, gem31Pro, and gem31FlashLite only surface when preview routing is explicitly enabled.
  • mimoV2Flash and mistralSmall4 are raw-only eval models, not active defaults.
  • Raw model additions should be checked into the manifest explicitly.
  • Use npm run validate:models to compare the static manifest against OpenRouter before releasing roster changes.