Choosing the Best LLMs for OpenClaw in 2026

Pick a main model for tool-heavy work, a cheap model for triage, and optionally a local model for private summaries. Exact IDs change; confirm the string your installed OpenClaw lists before you paste a config.

November 2026: official notes through v2026.9.9 name Claude Opus 5.5, Claude Haiku 5.5, GPT-6 Sol / Luna / Astra, GPT-6.1 Sol, and Grok 4.7. Provider setup: overview. Wiring details: autumn 2026 models.

Operator matrix

FamilyUse in OpenClawWatch-outs
Claude Opus 5.5Default for long-context tool use, skill authoring, and specialist-team leads.Cost. Cap spend. Setup: Anthropic.
Claude Haiku 5.5Fast channel replies, classification, first-pass triage.Do not give it unrestricted shell as the only model.
GPT-6 Sol / Luna / AstraSol for general agent work; Luna where notes mention it for chat; Astra as a September option. GPT-6.1 Sol in Codex (v2026.9.9).Policy and billing live on the OpenAI account. OpenAI setup. Sign-in beta is a separate topic.
Grok 4.7Named in v2026.9.6 chat-model support.Confirm tool-calling on your OpenClaw build before making it default.
GeminiStill a reasonable cheap multimodal path.Verify tool schemas. Google.
DeepSeekCost-efficient reasoning.Latency and region. DeepSeek.
Ollama (local)Privacy, flat electricity cost, summarizers.Hardware. local guide.

Stacks that still make sense

  • Personal assistant: Opus 5.5 or GPT-6 Sol as main, Haiku or a mini GPT as fast, optional Ollama for private notes.
  • DevOps / code: Opus 5.5 or GPT-6 Sol with a strict sandbox. Pair with PR review.
  • High-volume chat: Haiku 5.5 or a fast GPT for first reply, escalate to Opus/Sol when tools run.
  • Offline lab: Ollama primary, cloud fallback only when the host is online.

Illustrative routing

This is a pattern, not a paste-and-forget schema. Your keys may differ.

models:
  default: anthropic/claude-opus-5.5
  fast: anthropic/claude-haiku-5.5
  coding: openai/gpt-6.1-sol
  local: ollama/llama
routing:
  channelReply: fast
  toolHeavy: default
  repoWork: coding
  summarize: local

Cost: cost playbook and 30-day Usage. Errors: model troubleshooting.

Related