Best model for OpenClaw (2026): price, tool use and safety
OpenClaw's docs say to use the strongest model you can for agents with tools. What that costs, which models top its list, what users route and how to switch.
OpenClaw’s documentation answers this directly: “Use the strongest latest-generation model your provider stack offers, especially for tool-enabled or untrusted-input agents”, because weaker models are easier to hijack (FAQ). As of October 2026, the top of OpenClaw’s own recommended list is Claude Opus 5.5, at $4 per million input tokens and $20 per million output, or about $290 a month at moderate use with OpenClaw’s default 30-minute heartbeat; GPT-6.1 Sol, second on the list, costs half that. A cheap model is fine in the case the docs carve out: “chat-only personal assistants with trusted input and no tools”.
We haven’t run OpenClaw or compared models inside it ourselves. This guide rests on OpenClaw’s documentation and model list as read on October 20, 2026, OpenRouter’s public usage data, and API prices we checked at each provider on October 17 and 18 (see how we work).
What OpenClaw’s docs recommend, and why
The FAQ line above isn’t a one-off. The same answer adds “Route cheaper models to routine/low-stakes chat by agent role”, and two other pages agree:
- The models page: “Set your primary to the strongest latest-generation model available to you” and “For tool-enabled agents or untrusted inputs, avoid older/weaker model tiers” (models).
- The security page: “Use the latest-generation, best-tier model for any bot that can run tools or touch files/networks” (prompt injection).
The reason is prompt injection: an agent that reads email, web pages or group chats reads text written by others, which can contain instructions aimed at the model. Per the FAQ, “weaker or over-quantized models are more vulnerable to prompt injection and unsafe behavior.”
The numbers the docs cite
OpenClaw’s security page cites “a 2026 crowdsourced arena of 272K attacks across 41 agent scenarios”, counting an attack only when the agent both executed the harmful action and hid it from the user. Success rates were 0.5% for Claude Opus 4.5, 1.0% for Sonnet 4.5, 1.3% for Haiku 4.5 and 8.5% for Gemini 2.5 Pro. The page doesn’t name or link the study.
Three things are worth taking from those figures:
- The vendor gap was bigger than the size gap. Opus, Sonnet and Haiku ranged from 0.5% to 1.3%; Gemini 2.5 Pro was at 8.5%. “Strongest” isn’t only about the biggest tier.
- They are older models. None of the four is in our current price data, so treat the study as a direction, not a ranking of today’s models.
- No model is enough on its own. The same page says adaptive human attackers reach “success rates above 80% against state-of-the-art defenses once the attacker adapts”, and calls model choice “your first and cheapest layer”. Sandboxing, tool allowlists and pairing still matter; see is OpenClaw safe?
When a cheaper model is fine
The security page’s exception: “For chat-only personal assistants with trusted input and no tools, smaller models are usually fine.” If a small model must have tools, it says to use “read-only tools, strong sandboxing, minimal filesystem access, strict allowlists.” That rules out a budget model for an assistant that reads your inbox or browses the web on its own, unless you accept those limits.
OpenClaw’s recommended models list
OpenClaw keeps a ranked list in its repository, scripts/lib/recommended-models.json: “a JSON array of normalized model ids, best first.” The order starts from each model’s Arena agent score, with ties broken and gaps filled by observed usage, and “the reviewer’s edit is the source of truth” (recommended models). OpenClaw’s model pickers show a provider’s recommended models first. The file holds 30 ids with no reasons attached. Here are the first ten, as read on October 20, 2026 (GitHub), with prices from our model price data:
| # | Id in OpenClaw’s list | Input / output per 1M tokens | Note |
|---|---|---|---|
| 1 | claude-opus-5.5 | $4 / $20 | Default for a new Anthropic setup |
| 2 | gpt-6.1-sol | $2 / $10 | |
| 3 | gemini-4-argon | $2 / $10 | Rises to $4 / $20 on December 1, 2026 |
| 4 | gpt-6-astra | $10 / $50 | OpenClaw’s default for a fresh setup |
| 5 | claude-fable-5.1 | $10 / $50 | Anthropic’s top model |
| 6 | claude-sonnet-5.5 | $2 / $10 | |
| 7 | muse-wonder-1 | $4 / $20 | Meta’s Muse Wonder, in preview |
| 8 | muse-wonder | not in our data | |
| 9 | claude-sonnet-5 | not in our data | |
| 10 | grok-4.7 | $2 / $6 |
Two things stand out. First, the list isn’t sorted by price. Two of the top three cost $2 / $10, while both $10 / $50 models sit fourth and fifth: by OpenClaw’s own ranking, paying five times more per token doesn’t buy a higher place. Second, there’s no budget model in the top ten. Smaller, local-sized models such as qwen3.8-27b, glm-5.3-flash and deepseek-v4-flash appear further down.
One caution: the docs don’t say the ranking reflects resistance to prompt injection.
The defaults, and what they cost
“A fresh setup with no primary model uses exact openai/gpt-6-astra” (FAQ), and a fresh Anthropic setup “defaults to Opus 5.5” (Anthropic provider). Through OpenRouter, “Onboarding defaults to openrouter/auto” (OpenRouter provider).
GPT-6 Astra costs $10 / $50 per million tokens, with cached input at $1. At the calculator’s default usage, below, that’s about $742 a month for the model alone. Opus 5.5, first on OpenClaw’s list, comes to about $290, and GPT-6.1 Sol, second, to about $145. So moving from the fresh-setup default to the list’s top pick cuts the model bill by about 60%, and to its second pick by about 80%, for the same usage.
One note from our price data: Claude Opus 5.5 uses a newer tokenizer that produces roughly 30% more tokens for the same text than Claude 4.6 and earlier, so token counts aren’t comparable across Claude generations.
What each model costs at your usage
The calculator’s defaults come from OpenClaw’s docs. The /context example shows a system prompt of about 9.6K tokens plus about 8.0K tokens of tool schemas before any conversation, about 17.6K in all; the calculator rounds up to 30,000 input tokens a call to allow for history. The other defaults: 600 output tokens a call, 3 calls per message (a reply that uses tools takes several), 20 messages a day, a heartbeat every 30 minutes around the clock at 30,000 tokens, and 60% of message input served from the prompt cache. For the full breakdown, see how much OpenClaw costs to run.
The server is the small part. As of October 2026, a Hetzner CX23 costs $6.49 a month, a Hostinger KVM 2 costs $8.99 a month at the intro price (renewing at $14.99), and a DigitalOcean 2 GiB Droplet costs $12. More plans are in VPS and app hosting.
What each model costs a month in OpenClaw
Pick a model and a machine. The table below the result runs the same usage through every model in our price data, cheapest first. Change the token assumptions, or set a cheaper heartbeat model, under “Token assumptions and heartbeat”.
Token assumptions and heartbeat
| Server | ||
|---|---|---|
| Your messages: input | ||
| Your messages: output | ||
| Heartbeat |
The same usage on every model
| Model | Per message | Model, a month | With server |
|---|
Model prices are list API prices per million tokens (input / output) as last checked 2026-10-18; server prices from our hosting comparison. Months are 30 days. Cache writes (Anthropic charges 1.25 times the input price for a 5-minute cache write), taxes and any free tiers aren't counted. An estimate from your inputs, not a measured bill.
At these settings, the models fall into clear bands for the model alone: under $25 a month for Claude Haiku 5.5, GPT-6 Luna, Mistral Small 4 and DeepSeek V4.1 Flash; about $56 for GPT-5.4 mini and Gemini 3.8 Flash; $145 to $155 for Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon and Grok 4.7; about $290 for Opus 5.5 and Muse Wonder; and over $700 for Claude Fable 5.1 and GPT-6 Astra.
What OpenClaw users actually run on OpenRouter
OpenRouter publishes which models each app sends traffic to. On OpenClaw’s app page, the top models this month were Step 5 Preview (1.66 trillion tokens), DeepSeek V4.1 Flash (1.29 trillion), GLM 5.3 Flash (666 billion), and two DeepSeek V4 Flash versions (416 and 341 billion), out of 502 models used (OpenRouter).
None of those is in the top ten of OpenClaw’s list, and the one in our price data is cheap: DeepSeek V4.1 Flash costs $0.30 / $1.20 per million tokens on DeepSeek’s own API at peak hours, about $21 a month at the calculator’s default usage, against about $290 for Opus 5.5.
Read this with care: it counts only traffic OpenClaw sends through OpenRouter (OpenRouter provider), not people who connect Anthropic or OpenAI directly or use a subscription, and it’s weighted by tokens, so heavy users count for a lot. Still, it suggests many OpenClaw users on OpenRouter choose on price, against the docs’ advice. If you do the same for an agent with tools, the docs’ safeguards (sandboxing, read-only tools, allowlists) stop being optional.
Model tiers and prices
The table lists every model in our price data. The chips group them by facts, not quality scores we don’t have: OpenClaw’s top ten, price band, open weights, and prices about to change.
Models for OpenClaw: prices per million tokens
Input, output and cached-input prices in US dollars per 1M tokens, and the context window, from our price data. Filter by group.
| Model | Input | Output | Cached | Context | Note |
|---|---|---|---|---|---|
| Claude Fable 5.1 (Anthropic) | $10 | $50 | $0.25 | 1M | No. 5 on OpenClaw's list |
| GPT-6 Astra (OpenAI) | $10 | $50 | $1 | 1.05M | OpenClaw's fresh-setup default; No. 4. Prompts over 272K tokens: $20 / $75 |
| Claude Opus 5.5 (Anthropic) | $4 | $20 | $0.20 | 1M | No. 1; Anthropic setup default |
| Muse Wonder (Meta) | $4 | $20 | $0.20 | 1.05M | No. 7 (as muse-wonder-1); preview |
| GPT-6.1 Sol (OpenAI) | $2 | $10 | $0.10 | 1.05M | No. 2 |
| Gemini 4 Argon (Google) | $2 | $10 | $0.10 | 1.05M | No. 3; $4 / $20 from Dec 1, 2026 |
| Claude Sonnet 5.5 (Anthropic) | $2 | $10 | $0.10 | 1M | No. 6 |
| Grok 4.7 (xAI) | $2 | $6 | $0.50 | 500K | No. 10. Prompts of 200K+ tokens: $4 / $12 |
| Gemini 3.1 Pro preview (Google) | $2 | $12 | $0.20 | 1.05M | Prompts over 200K tokens: $4 / $18 |
| Mistral Medium 3.5 | $1.50 | $7.50 | — | 256K | No cached price listed |
| DeepSeek V4 Pro | $1.32 | $3.96 | $0.044 | 1M | Peak-hour price; half off-peak |
| Grok 4.3 (xAI) | $1.25 | $2.50 | $0.20 | 1M | Older Grok |
| Muse Spark 1.3 (Meta) | $1.25 | $4.25 | $0.15 | 1.05M | Standard mode |
| Llama 3.3 70B (Together AI) | $1.04 | $1.04 | — | 128K | Open-weight; Together's price |
| Mistral Large 4 | $0.68 | $2.09 | $0.07 | 1M | Launch sale; list price $1.36 / $4.18 expected from about Oct 20 |
| GPT-5.4 mini (OpenAI) | $0.75 | $4.50 | $0.075 | 400K | Previous generation |
| Gemini 3.8 Flash (Google) | $0.75 | $3.75 | $0.075 | 1.05M | Doubles from Jan 1, 2027 |
| DeepSeek V4.1 Flash | $0.30 | $1.20 | $0.006 | 1M | Peak-hour price; No. 2 for OpenClaw on OpenRouter |
| Gemini 3.5 Flash-Lite (Google) | $0.30 | $2.50 | $0.03 | 1.05M | |
| GPT-5.4 nano (OpenAI) | $0.20 | $1.25 | $0.02 | 400K | Deprecated; leaves the API Apr 1, 2027 |
| Mistral Small 4 | $0.15 | $0.60 | — | 256K | Apache 2.0 |
| GPT-OSS 120B (Together AI) | $0.15 | $0.60 | — | 131K | OpenAI's open-weight model; Together's price |
| Claude Haiku 5.5 (Anthropic) | $0.10 | $0.50 | $0.01 | 1M | Prompts over 100K tokens: $0.50 / $2.50 |
| GPT-6 Luna (OpenAI) | $0.10 | $0.50 | $0.01 | 1.05M | Smallest GPT-6 |
Showing all 24.
Prices checked at each provider on October 17–18, 2026 (all model prices). "No." is the position in OpenClaw's recommended-models.json on October 20, 2026. Two of the top ten, muse-wonder and claude-sonnet-5, are not in our price data. Budget: under $1 per 1M input tokens. Mid-price: $1 to $2, not in the top 10 (Mistral Large 4 by its list price). Context: 1.05M stands for 1,048,576 or 1,050,000 tokens.
Context isn’t the constraint for API models: the smallest window here is 128K tokens, against the roughly 17.6K OpenClaw sends before any history. Watch the long-prompt rates instead. Claude Haiku 5.5 costs five times as much past 100,000 tokens, a size OpenClaw’s docs say a long main session can reach (heartbeat).
Using a subscription instead of an API key
A plan you already pay for can stand in for per-token billing, within its limits.
OpenAI: Codex and Sign in with ChatGPT
OpenClaw supports three OpenAI logins: a Codex login that uses your ChatGPT account, “Sign in with ChatGPT (Beta)”, which lets OpenClaw use your Codex allowance, and a Platform API key (OpenAI authentication). You pick one with openclaw models auth login --provider openai --method oauth|device-code|siwc|api-key. When both a plan and a key are set up, “automatic selection prefers the subscription route”, and when you hit the plan’s limit OpenClaw blocks that login until the reset (OpenAI runtimes).
The catches: Sign in with ChatGPT doesn’t support realtime voice or image generation in OpenClaw. If your Codex workspace lacks access to the default model, OpenClaw “does not silently downgrade”, so you may have to pick a model yourself (FAQ). On OpenAI’s side, its directory lists OpenClaw under “Open-source integrations” (OpenAI).
Anthropic: possible, with caveats
OpenClaw can reuse your Claude Code login or a token from claude setup-token. Its docs call Claude CLI reuse “a sanctioned auth path”, based on what Anthropic staff told the project, but also say “For Anthropic in production, API key auth is still the safer recommended path” and warn that “Anthropic token auth expires and can be revoked” (OAuth, Anthropic provider).
Anthropic’s own position has shifted during 2026. Subscriptions stopped covering third-party tools on April 4; an October 7 Help Center update says you “can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits” (Anthropic Help Center); and the Claude Code terms still bar third-party developers from routing requests through Free, Pro or Max credentials on their users’ behalf (Claude Code legal). Our history of OpenClaw has the timeline.
Two practical effects: with Anthropic OAuth or token auth, the heartbeat default slows from every 30 minutes to every hour, and /usage can’t show a dollar cost for Claude CLI usage (heartbeat, usage costs).
A cheaper model for heartbeats, compaction and side jobs
Side jobs can run on a cheaper model than the one that handles your requests:
- Heartbeat.
agents.defaults.heartbeat.modelsets a model for heartbeat runs.isolatedSession: truedrops the conversation history (“~100K tokens down to ~2-5K per run”), andlightContext: trueskips the workspace files (heartbeat). - Compaction.
agents.defaults.compaction.modelsets the model that summarizes long sessions (compaction). - Utility tasks.
agents.defaults.utilityModelis “a lower-cost model for short internal tasks” (models).
At the calculator’s defaults, Opus 5.5 costs about $290 a month, of which about $176 is the heartbeat. Move only the heartbeat to Claude Haiku 5.5 and the total falls to about $119 (set “Heartbeat model” in the calculator to check).
One caution, which is our reading rather than the docs’: a heartbeat runs a full agent turn with your tools. If it checks your inbox or the web, it reads untrusted input with tools at hand, which is the case where the docs warn against weak models. A cheap heartbeat model fits best when the check is narrow. A setup along these lines, using model names from OpenClaw’s Anthropic page:
{
agents: {
defaults: {
model: {
primary: "anthropic/claude-opus-5-5",
fallbacks: ["anthropic/claude-sonnet-5-5"],
},
heartbeat: {
every: "1h",
model: "anthropic/claude-haiku-5-5",
isolatedSession: true,
lightContext: true,
},
},
},
}How to change the model
- In chat:
/model <model>changes the current session. Add-ato also make it the agent’s default, or-gthe shared global default;/model default -sclears a session pin (models). - From the command line:
openclaw models set <provider/model>sets it;openclaw models listandopenclaw models statusshow what’s available and the state of each login. - In the config file:
agents.defaults.model.primaryand the orderedfallbackslist; per-agent overrides go underagents.entries.*.model, andagents.defaults.modelPolicy.allowlimits which models can be chosen.
Models are named provider/model, such as anthropic/claude-opus-5-5; OpenRouter models use openrouter/<provider>/<model>. When a model fails, OpenClaw first rotates logins at the same provider, then moves to the next fallback. A model you pick explicitly with /model is strict, though: if it fails before replying, OpenClaw reports the failure instead of falling back (model failover).
Local models, briefly
Running the model yourself removes the token bill, not the docs’ warnings. They advise the “largest available model build” you can run, warn that “A model that loads or answers a short prompt may still fail an agent turn”, and note that “Local models do not provide hosted providers’ safety filters” (local models). Ollama and LM Studio models need tool support and at least 16K of context for OpenClaw to pick them up, while Ollama itself recommends at least 64K for agents. Our guide to OpenClaw with local models runs the numbers, and how to rent a GPU covers the cloud option.
Which model to pick
- An agent with tools that reads email, the web or group chats: a model from the top of OpenClaw’s list. Claude Opus 5.5 is first; GPT-6.1 Sol (second) and Claude Sonnet 5.5 (sixth) cost half as much per token.
- You already pay for ChatGPT: sign in with it, and set an API model as a fallback for when you hit the plan’s limit.
- A chat-only assistant with trusted input and no tools: a budget model such as Claude Haiku 5.5, GPT-6 Luna or DeepSeek V4.1 Flash.
- In every case: put the heartbeat, compaction and utility jobs on a cheaper model, and turn on sandboxing. Compare other agents in self-hosted AI agents, and model profiles in AI models.
FAQ
What is the best model for OpenClaw?
Per OpenClaw’s docs, the strongest latest-generation model you have access to, especially if the agent uses tools or reads untrusted content. As of October 20, 2026, OpenClaw’s recommended list starts with Claude Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon.
What model does OpenClaw use by default?
A fresh setup with no primary model uses openai/gpt-6-astra, which costs $10 / $50 per million tokens. A fresh Anthropic setup uses Claude Opus 5.5, and OpenRouter onboarding uses openrouter/auto.
Can I use OpenClaw with Claude?
Yes: with an Anthropic API key, which OpenClaw’s docs recommend for production, or by reusing your Claude Code login or a setup token. Anthropic’s rules on subscriptions in third-party tools changed several times in 2026, so check its current terms first.
Can I use OpenClaw with Codex or a ChatGPT plan?
Yes. OpenClaw supports a Codex login and “Sign in with ChatGPT (Beta)”, which draw on your plan’s Codex allowance instead of API billing, until you reach the plan’s limit.
How do I change the model in OpenClaw?
Type /model <provider/model> in chat to switch the current session (add -a or -g to make it a default), or run openclaw models set <provider/model>. Fallbacks go in agents.defaults.model.fallbacks.