OpenClaw LLM Compatibility: Choosing Between Claude, GPT-4, DeepSeek, and Local Models product guide

Now I have comprehensive, well-sourced material to write the article. Let me compose the verified, authoritative cluster article.

Why LLM Selection Is the Most Consequential Decision in Your OpenClaw Setup

Most OpenClaw guides treat model selection as a footnote — a single line in a configuration file you fill in once and forget. That framing is a mistake. The LLM you wire into OpenClaw determines not just the quality of every task your agent executes, but also your monthly operating cost, your response latency, and — for Australian businesses in particular — whether your deployment is legally compliant with the Privacy Act 1988 and the Australian Privacy Principles.

Unlike closed platforms run by a single vendor such as ChatGPT or Claude.ai, OpenClaw's standout feature is its multi-model architecture — you can freely switch between Anthropic's Claude, OpenAI's GPT, Google's Gemini, DeepSeek's open-source models, and even locally deployed Llama models via Ollama, all managed through the unified provider/model configuration format (e.g., anthropic/claude-sonnet-4-6).

That flexibility is powerful, but it creates a decision surface that deserves genuine strategic analysis. This article provides exactly that.

What Models Does OpenClaw Support?

OpenClaw supports 15+ language models from leading AI providers, including cloud APIs like DeepSeek V3, OpenAI GPT-4o, and Anthropic Claude 3.5 Sonnet, as well as local models running for free through Ollama.

The model providers page covers LLM and model providers specifically (not chat channels like WhatsApp/Telegram). Model refs use the provider/model format (example: anthropic/claude-opus-4-6).

The full current roster of supported providers includes:

• Anthropic — Claude Opus, Sonnet, and Haiku families
• OpenAI — GPT-4o, GPT-4, and forward-compatible GPT-5.x variants
• Google — Gemini Pro and Gemini Flash families
• DeepSeek — DeepSeek V3, V3.2, and R1
• Moonshot AI (Kimi) — Kimi K2, K2.5
• MiniMax — M2.5 and M2.5 Lightning
• Zhipu AI — GLM-4.x and GLM-5.x families
• Qwen — Qwen3, Qwen3 Coder, Qwen 2.5 Coder families
• Local inference — Any OpenAI-compatible endpoint via Ollama, vLLM, LM Studio, or llama.cpp
• OpenRouter — Unified API routing to most of the above

OpenClaw supports any LLM that exposes an OpenAI-compatible API endpoint, including Ollama, vLLM, LM Studio, and llama.cpp.

The Four Primary Dimensions of Model Selection

Model selection is critical to the OpenClaw experience because different models vary significantly in reasoning capability, response speed, tool-calling reliability, and price.

These dimensions compete with each other. You can optimise for two, maybe, but rarely all three simultaneously.

1. Capability: Tool Calling Is the Non-Negotiable Baseline

In a standard chatbot, model quality maps roughly to response coherence. In OpenClaw's agentic loop, the relevant capability is tool-calling reliability — the model's ability to invoke shell commands, write to files, call external APIs, and chain multi-step actions without malformed outputs or dropped context.

Some models — particularly smaller open-source models — do not support function calling, which OpenClaw relies on for agent actions like file editing and command execution. The fix is to switch to a model that supports tool calling.

Benchmarks can be misleading. For OpenClaw, what actually matters is: tool calling — can it invoke shell commands and APIs without fumbling the syntax? — and context tracking — does it remember what you said 50 messages ago?

For OpenClaw specifically, Claude models outperform GPT-4o on long-context tasks, prompt-injection resistance, and multi-step tool use.

2. Cost: The Range Is Staggering

The price spread across supported models is staggering: the most expensive model, Claude Opus 4.6, has an output price nearly 60× that of the cheapest, DeepSeek V3.2.

Approximate monthly API costs based on real user patterns: top-tier Claude (Opus level): light use $80–150, moderate $200–400, heavy $500–750+; middle-tier Claude (Sonnet level): light $15–30, moderate $40–80, heavy $100–200; GPT-4o: light $12–25, moderate $30–60, heavy $80–150; DeepSeek or light Claude (Haiku level): light to moderate $5–15 or less, heavy usually under $30; local models via Ollama: $0 API cost (only electricity and hardware).

3. Latency: Cloud vs. Local Trade-offs

Cloud APIs from Anthropic and OpenAI deliver consistently fast responses because they run on purpose-built inference infrastructure. Local models introduce latency that scales with model size and hardware.

A 7B model needs approximately 6GB VRAM. A 32B model quantised to Q4 needs approximately 20GB. Most consumer GPUs (RTX 3090, RTX 4090) handle 32B models well. Apple Silicon Macs with 32GB+ unified memory can run 32B models and even some MoE models.

Without a GPU, CPU inference with llama.cpp is possible — but expect 10–20× slower responses.

4. Data Sovereignty: The Dimension Competitors Ignore

For Australian users and businesses, data sovereignty is not an abstract concern — it is a legal obligation. This dimension is addressed in full in the section below.

Provider-by-Provider Analysis

Anthropic Claude: The Default Recommendation

Claude Opus 4.6 is the best AI model for OpenClaw in 2026 if quality and safety matter most. OpenClaw's creator, Peter Steinberger, explicitly recommends Anthropic Pro/Max with Opus 4.6 for its long-context strength and superior prompt-injection resistance.

For most OpenClaw users, Claude Sonnet 4.5 hits the sweet spot between capability and cost. It handles email automation, calendar management, and web browsing at a fraction of Opus pricing.

Claude has become the default for coding agents. The tool use is simply more reliable than the alternatives.

Configuring Claude in OpenClaw:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-opus-4-6"
      }
    }
  }
}

Direct public Anthropic requests support the shared /fast toggle and params.fastMode. OpenClaw now prefers Claude CLI reuse and claude -p when available, with setup-token remaining available as a supported token path.

Key considerations for Anthropic:

• Data is processed on Anthropic's US-based servers
• Anthropic's API terms prohibit training on API data by default
• Best-in-class prompt-injection resistance — critical for agentic deployments (see our guide on OpenClaw Security Risks)
• Cost premium is justified for high-stakes workflows involving financial decisions, client communications, or sensitive data

OpenAI GPT-4o: The Versatile Runner-Up

GPT-4o slots between Sonnet and Haiku on price. For OpenClaw specifically, Sonnet 4.5 outperforms it on agent-specific tasks. But if you already have OpenAI credits, GPT-4o is a solid option.

GPT-4o offers solid general performance and fast responses, making it a good all-rounder.

On the cloud side, Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro all have robust tool calling support.

GPT-4o's primary advantage over Claude is speed on simpler tasks and its deep ecosystem of community documentation. Its disadvantage in the OpenClaw context is slightly weaker multi-step tool reliability on complex, long-context agentic chains.

Google Gemini: The Third Option

OpenClaw supports Google as a first-party provider. Gemini's principal advantage is its extended context window and competitive pricing on the Flash tier. Google's open-source Gemma 4 models also perform well on instruction-following tasks and are worth benchmarking against Qwen for specific use cases.

DeepSeek: Powerful for Cost Optimisation, Problematic for Australian Compliance

DeepSeek's pricing is its headline advantage.

DeepSeek V3 has become the go-to budget model for OpenClaw users who want to minimise costs. At roughly $0.27 per million input tokens and $1.10 per million output tokens, it's 10–50× cheaper than Claude Opus.

Many users run DeepSeek for 80–90% of routine work and switch to stronger models only for sensitive or complex cases — this keeps bills very low while covering what matters.

However, DeepSeek's capability limitations are real. DeepSeek handles basic OpenClaw tasks reasonably well — simple commands, routine automation, basic email processing — but it struggles with complex multi-step reasoning and long context windows, and prompt-injection resistance is significantly weaker than Claude or GPT-4o.

For Australian users, the compliance picture is unambiguous and severe. The Australian Department of Home Affairs issued a federal directive on February 4, 2025, noting that DeepSeek "poses an unacceptable level of security risk" due to its susceptibility to jailbreak attacks and data transmission concerns.

Under the directive, Australian government agencies must prohibit the installation or use of DeepSeek software and services and immediately remove existing instances. The sweeping bans were imposed by Queensland, Western Australia, the ACT, South Australia, New South Wales, and the Northern Territory.

DeepSeek stores all user data on servers located in the People's Republic of China. This isn't merely a technical detail — it has profound legal implications. Under Chinese intelligence laws, particularly the 2017 National Intelligence Law, organisations and individuals must "support, assist, and cooperate with national intelligence efforts." This means Chinese authorities can legally compel DeepSeek to hand over user data upon request, with no requirement to notify affected users.

The takeaway for Australian businesses is clear: using the DeepSeek API routes your agent's prompts — and any data embedded in them — through China-based servers. Using the DeepSeek open-source model weights locally via Ollama is an entirely different matter, and is discussed below.

Local Models: The Data Sovereignty Solution

Running OpenClaw against a locally hosted model is the only approach that provides complete data sovereignty. If you want zero chance of your data ever leaving your machine, local models via Ollama are the only real way. You download an open-weight model, run it on your own hardware, and OpenClaw talks to it directly through a local server — no API keys, no cloud provider, no logs anywhere else.

This matters directly under Australian law. The Privacy Act 1988 and the Australian Privacy Principles (APPs) apply to all uses of AI involving personal information, including where information is used to train, test, or use an AI system.

Data sovereignty means your customer data remains subject to Australian law because it's stored on servers physically located in Australia. This matters because the US CLOUD Act allows American law enforcement to access data stored by US companies, regardless of where those companies operate. If your AI service uses US-based infrastructure, Australian customer data may be accessible to foreign governments without your knowledge.

The first tranche of Privacy Act reforms, passed in 2024, introduced new transparency obligations around automated decision-making that will take effect in December 2026. Australia's privacy regulator, the Office of the Australian Information Commissioner, has been proactive in interpreting the Act in AI contexts and is actively regulating AI through interpretation and enforcement rather than waiting for dedicated legislation.

Hardware Requirements for Local Inference

For most setups, Qwen 3 32B (20GB VRAM) offers the best balance of reasoning, tool calling, and speed. For coding-focused agents, Qwen 2.5 Coder 14B (10GB VRAM) is the sweet spot. On limited hardware (8GB), Gemma 3 9B is the best option. For enterprise setups with 128GB+ unified memory, Qwen 3.5 (397B MoE) and MiniMax M2.5 deliver near-Claude-level performance locally.

7–8B models (Qwen 3 8B, Mistral 7B, Llama 3.2 8B) produce tool call format errors constantly. They are fine for answering one-off questions, but are effectively useless for agent work.

Practical hardware tiers for local OpenClaw deployment:

32GB unified memory is where local-first actually starts working. OpenClaw (~1GB) + Devstral-24B at Q4_K_M (~14GB) + KV cache at 65K (~4–6GB) + OS (~3GB) = about 24GB total — fits with headroom.

Configuring Ollama in OpenClaw

The most common local model configuration failure is a missing api: "openai-responses" in the Ollama config. The correct configuration is:

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://127.0.0.1:11434/v1",
        "apiKey": "ollama-local",
        "api": "openai-responses"
      }
    }
  }
}

The Multi-Model Strategy: Cost Optimisation Without Capability Sacrifice

OpenClaw's fallback chain configuration enables a tiered routing strategy that dramatically reduces costs without degrading output quality for most interactions.

When OpenClaw asks a clarifying question, generates a simple file, or executes a routine command, a $0.40/M token model performs just as well as Claude Opus at $75/M output tokens. By routing simple tasks to cheap models and reserving expensive models for complex reasoning, you dramatically reduce your average cost per interaction without meaningfully impacting quality. OpenClaw supports this through its fallback chain configuration in agents.defaults.model.fallbacks.

Consider a typical developer using OpenClaw for four hours of active coding per day. With Claude Sonnet 4.5 as the only model, that is roughly 2–4 million tokens per day at a blended average cost of around $9/M tokens, resulting in $18–36 per day or $400–800 per month. With multi-model routing — using DeepSeek V3.2 at $0.53/M for 70% of interactions and Claude Sonnet for the remaining 30% — the same usage pattern drops to approximately $5–12 per day or $100–260 per month. That is a 55–67% reduction in costs with minimal impact on output quality, because the majority of agent interactions (file reads, simple edits, command execution, status checks) do not benefit from a premium model's reasoning capabilities.

Recommended three-tier configuration:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-6",
        "fallbacks": [
          "openrouter/deepseek/deepseek-v3.2",
          "ollama/qwen3-coder:32b"
        ]
      }
    }
  }
}

This pattern uses Claude Sonnet as the primary for reliable tool use, routes to DeepSeek V3.2 for cost relief on routine tasks, and falls back to a local Ollama model if both cloud providers are unavailable — ensuring the agent never stops working entirely. For Australian businesses handling personal data, the DeepSeek fallback should be replaced with a second Claude tier or a local model.

Model Switching: Configuration and Common Failure Modes

Model switching issues typically stem from: (1) missing required configuration fields like baseUrl for custom providers, (2) the Gateway not being restarted after config changes, (3) stale session state caching the old model, (4) Docker networking issues preventing container-to-container communication, (5) incorrect model name format, or (6) provider API compatibility settings missing (like api: "openai-responses" for Ollama).

Troubleshooting sequence for a model that won't switch:

1. openclaw gateway restart — resolves the majority of cases
2. /new in the chat interface — clears cached session state
3. openclaw doctor --fix — validates configuration
4. openclaw models list — confirms which models are currently available

Configuring a model in OpenClaw takes a single command. Set your preferred model and API key, then OpenClaw handles all API communication, token management, and error handling automatically.

Australian Data Sovereignty: A Decision Framework

The Australian compliance picture for LLM selection can be summarised as follows:

While the Australian government has not specifically addressed agentic AI in any released policies, guidelines, or laws, the broader principles for responsible and trustworthy AI are likely to be applied to the development of agentic AI as well. The 2024 amendments to the Privacy Act, which will come into effect in late 2026, have significant ramifications for automated decision-making. Covered entities must now disclose, within their privacy policies, the types of personal information used, the nature of decisions made solely by computer programs, and those where computer assistance significantly influences outcomes that could substantially affect individuals' rights or interests.

With maximum penalties reaching $50 million, three times the benefit obtained, or 30% of adjusted turnover, organisations face financial exposure that rivals or exceeds many of the world's most stringent privacy frameworks.

For Australian healthcare, legal, and financial services organisations — sectors where OpenClaw's document automation and calendar management capabilities are particularly valuable — the practical recommendation is: use Claude or GPT-4o with explicit data processing agreements, or run local inference on Australian-hosted hardware. The managed hosting options discussed in our companion guide (OpenClaw Managed Hosting in Australia: Data Sovereignty, Compliance, and Provider Options) provide a third path that combines local inference with managed operations.

Key Takeaways

• OpenClaw's multi-model architecture allows you to freely switch between Anthropic's Claude, OpenAI's GPT, Google's Gemini, DeepSeek, and locally deployed models via Ollama, all managed through the unified provider/model configuration format.

• Claude is the recommended default for most OpenClaw users. OpenClaw's creator Peter Steinberger explicitly recommends Anthropic Pro/Max with Opus 4.6 for its long-context strength and superior prompt-injection resistance, while Claude Sonnet is the practical sweet spot for cost-conscious deployments.

• DeepSeek's API is incompatible with Australian compliance requirements. The Australian Department of Home Affairs issued a federal directive noting that DeepSeek "poses an unacceptable level of security risk." However, running DeepSeek's open-source model weights locally via Ollama eliminates this concern.

• Multi-model routing can reduce API costs by 55–67% by routing routine agent interactions to cheaper models while reserving premium models for complex reasoning tasks.

• Local inference via Ollama is the only approach that guarantees data never leaves your network. Models of 32B parameters or larger are required for reliable tool calling in production agentic workflows; smaller models produce tool call format errors that break the agent loop.

Conclusion

LLM selection in OpenClaw is not a one-time configuration decision — it is an ongoing strategic variable that determines your agent's capability ceiling, your operating costs, and your legal exposure. The model-agnostic architecture that makes OpenClaw so powerful also means that every deployment decision is yours to own.

For most users, the practical starting point is Claude Sonnet as the primary model, with a local Ollama fallback for resilience. Australian businesses handling personal data should treat local inference not as a performance compromise but as a compliance requirement, particularly as the Privacy Act reforms scheduled for late 2026 introduce new automated decision-making transparency obligations.

For a deeper understanding of how OpenClaw's Gateway routes messages to your chosen LLM, see our guide on How OpenClaw Works: The Gateway, Agent Loop, Skills System, and Memory Architecture. For the full deployment and security picture, including how model selection interacts with prompt injection risk, see OpenClaw Security Risks: Prompt Injection, Malicious Skills, and Safe Deployment Practices. And for Australian businesses evaluating managed hosting options that include local inference without third-party API calls, see OpenClaw Managed Hosting in Australia: Data Sovereignty, Compliance, and Provider Options.

References

• Office of the Australian Information Commissioner (OAIC). "Guidance on Privacy and the Use of Commercially Available AI Products." OAIC, 2025. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products

• International Association of Privacy Professionals (IAPP). "Global AI Governance Law and Policy: Australia." IAPP, 2026. https://iapp.org/resources/article/global-ai-governance-australia

• Reed Smith LLP. "Australia in Focus: Data Protection and AI in Australia." Reed Smith Viewpoints, February 2026. https://www.reedsmith.com/our-insights/blogs/viewpoints/102mk8i/australia-in-focus-data-protection-and-ai-in-australia/

• Australian Department of Home Affairs. Federal directive prohibiting use of DeepSeek products on Australian Government systems, February 4, 2025. Referenced in: Scimex. "Expert Reaction: Australian Federal Government Bans Chinese AI DeepSeek on Devices." Scimex, 2025. https://www.scimex.org/newsfeed/expert-reaction-australian-federal-government-bans-chinese-ai-deepseek-on-devices

• OpenClaw Project. "Model Providers." OpenClaw Official Documentation, 2026. https://docs.openclaw.ai/concepts/model-providers

• OpenClaw Project. "How to Change Your AI Model in OpenClaw — Claude, GPT, Gemini." GetOpenClaw.ai, 2026. https://www.getopenclaw.ai/help/switching-models-provider-config

• LaoZhang AI Blog. "OpenClaw Best Model Selection Guide: Claude vs GPT vs Gemini vs DeepSeek (2026)." LaoZhang.ai, March 2026. https://blog.laozhang.ai/en/posts/openclaw-best-model-selection-guide

• Wikipedia contributors. "OpenClaw." Wikipedia, The Free Encyclopedia, April 2026. https://en.wikipedia.org/wiki/OpenClaw

• Clawctl Blog. "OpenClaw with Local LLM: The Complete Guide (Ollama, vLLM, LM Studio)." Clawctl.com, April 2026. https://www.clawctl.com/blog/openclaw-local-llm-complete-guide

• SafeAI-Aus. "Current Legal Landscape for AI in Australia." SafeAI-Aus.org, January 2026. https://safeaiaus.org/safety-standards/ai-australian-legislation/