Best AI Model for OpenClaw (July 2026): Ranked on Real Use

Do not index

Five brass keys of different sizes held above a single lock, the metaphor for choosing an AI model for OpenClaw

The best AI model for OpenClaw is MiniMax M3 for most setups. It is what OpenClaw users actually run: 1.08 trillion tokens in 30 days, more than double the runner-up, at $0.30 per million input tokens. Reach for Claude Opus 5 when a task must not fail, and DeepSeek V4 Flash at $0.14 when the bill matters more than the last 20% of quality.

That answer is not a taste test. OpenRouter publishes what OpenClaw's 90 million monthly requests actually route to, and it separately publishes where money goes on agent work. Those two lists disagree with each other, and that disagreement is the useful part.

Every price, token count and benchmark figure below was pulled on 29 July 2026 and carries its source.

Best model for OpenClaw at a glance

If you want	Run this	Input / output per 1M	Why
The default most OpenClaw users land on	MiniMax M3	$0.30 / $1.20	The most-used model in real OpenClaw traffic, 1M context, handles images
The cheapest thing that still holds a tool loop	DeepSeek V4 Flash	$0.14 / $0.28	Second most-used, 1M context, roughly $8 a month for a personal assistant
A task that must not fail	Claude Opus 5	$5.00 / $25.00	Top of the agentic benchmark at 55.3, and where the money actually goes
The best score per dollar	GPT-5.6 Luna	$1.00 / $6.00	45.6 agentic at $0.67 a request against Opus 5's 55.3 at $12.61
Long sessions on a budget	DeepSeek V4 Pro	$0.44 / $0.87	Cache hits bill at $0.0036, under 1% of list
Zero budget	Nemotron 3 Super (free)	$0.00 / $0.00	Genuinely free with tool calling, but rate limited (see below)
Raw speed	gpt-oss-120b on Cerebras	$0.35 in	522 tokens per second
Long documents	Kimi K3	$3.00 / $15.00	1M context, cache hits at $0.30, 50.1 agentic
Everything private	See our best local LLM for OpenClaw guide	Free after hardware	A different decision with different limits

If you only read one line: run MiniMax M3 as your primary, point the utility model at something cheap, and switch to Claude with `/model` on the turns that matter. The rest of this guide is why, and how.

OpenClaw model ranking: what users actually run

OpenClaw is a publicly ranked app on OpenRouter. In the 30 days to 29 July 2026 it pushed 4.75 trillion tokens across 90.5 million requests, putting it in the same bracket as Claude Code and ahead of Cline, Cursor and LangChain.

OpenRouter breaks that traffic down by model on its Top AI Models Used by OpenClaw collection. It is the closest thing that exists to a census of what people run their agents on.

OpenRouter's OpenClaw collection, showing the top models by real 30-day token usage as of July 2026

#	Model	30-day tokens	Context	$/M in	$/M out
1	MiniMax M3	1.08T	1.05M	$0.30	$1.20
2	DeepSeek V4 Flash	556B	1.05M	$0.14	$0.28
3	Step 3.7 Flash	270B	262K	$0.20	$1.15
4	GLM 5.2	246B	1.05M	$0.69	$2.18
5	DeepSeek V4 Pro	223B	1.05M	$0.44	$0.87
6	Nemotron 3 Super (free tier)	222B	262K	$0.00	$0.00
7	Claude Sonnet 4.6	137B	1M	$3.00	$15.00
8	Claude Sonnet 5	133B	1M	$2.00	$10.00
9	Hy3 (Tencent)	133B	262K	$0.13	$0.53
10	Nemotron 3 Ultra (free tier)	115B	1M	$0.00	$0.00
11	Claude Opus 4.8	93B	1M	$5.00	$25.00
12	Claude Opus 4.6	77B	1M	$5.00	$25.00
13	Gemini 2.5 Flash Lite	67B	1.05M	$0.10	$0.40

That is the top 13 of the 20 models OpenRouter publishes, which together account for 3.72 trillion tokens, or 78% of OpenClaw's monthly total. Prices are the standard real-time rates from OpenRouter's model API. Percentages below are shares of that published top 20.

Three things fall out, and none of them match the advice you will find elsewhere.

Three quarters of the traffic goes to models under $0.50 per million input tokens. 74.4%, to be exact.

Every Claude model combined is 11.8%, which is about 9% of everything OpenClaw sent through OpenRouter. The common answer, "use Claude", describes roughly a tenth of what actually happens.

Nothing in the top five is American. MiniMax, DeepSeek, StepFun and Z.AI are the models carrying real agent workloads on price.

One honest caveat: this is OpenRouter traffic only. Anyone using a direct Anthropic or OpenAI key is invisible here, and those users skew expensive. Read it as "what the price-sensitive majority chose", not "what every OpenClaw user runs".

The models people run are not the models they pay for

OpenRouter tags every request by task and publishes spend share and token share side by side. The tag workflow execution is the single largest on the whole platform: 20% of all spend, 21.6% of all tokens. That is your OpenClaw agent's exact shape of work.

OpenRouter's spend by task, with workflow execution as the largest tag and Claude models leading it

Leaders by spend	Share	Leaders by tokens	Share
Claude Opus 4.8	19.6%	MiMo-V2.5	12.2%
Claude Opus 4.7	18.1%	DeepSeek V4 Flash	10.4%
GLM 5.2	8.1%	Hy3	9.6%
GPT-5.5	8.0%	MiniMax M3	8.8%
Claude Fable 5	6.4%	GLM 5.2	7.6%

Same task. Almost entirely different models. Claude models take more than half of the spend on agent workflow execution. The cheap mixture-of-experts models take the volume.

That is not a contradiction, it is the strategy people have converged on: run something cheap all day, and pay up on the small number of turns where being wrong is expensive. The section on configuration below shows how to actually do that, because the obvious way is the wrong way.

What an OpenClaw agent costs per month, by model

A chat message costs a fraction of a cent. An agent is a different animal, because every turn re-sends the system prompt, the tool definitions and the conversation so far. OpenClaw's docs are explicit that "everything the model receives counts, including: system prompt (all sections), conversation history, tool calls + tool results".

So here is the number nobody else on this topic will give you. A personal assistant on WhatsApp or Telegram: 40 messages a day, three model turns per message, 25,000 tokens of prompt per turn, 400 tokens out, 60% of input served from cache, plus the default half-hourly heartbeat running in isolated mode.

Model	Monthly model bill
Nemotron 3 Super (free tier)	$0, until you hit the rate limit
DeepSeek V4 Flash	about $8
MiniMax M3	about $18
GPT-5.6 Luna	about $58
Claude Sonnet 5	about $112
Claude Opus 5	about $279

Change the assumptions and every row moves, but the 36x spread between the cheapest and the most expensive stays. That spread, not any benchmark, is why the usage table looks the way it does. The two Claude rows are conservative, because Claude 4.7 and newer count roughly 30% more tokens for the same text.

Two mechanics drive it.

Your cached-input rate is your real price

After turn one, most of what you send is a repeat of what you already sent, so the rate that decides your bill is the cached read, not the headline input price.

Model	Input	Cached read	Cached as % of input
DeepSeek V4 Pro	$0.435	$0.0036	0.8%
Claude Opus 5	$5.00	$0.50	10%
Claude Sonnet 5	$2.00	$0.20	10%
GPT-5.6 Luna	$1.00	$0.10	10%
Kimi K3	$3.00	$0.30	10%
GLM 5.2	$0.69	$0.13	19%
MiniMax M3	$0.30	$0.06	20%
DeepSeek V4 Flash	$0.14	$0.028	20%

DeepSeek V4 Pro is the outlier by an order of magnitude. Its cache is on by default and needs no code changes, which makes it quietly the cheapest capable model on this list for a long-running session, despite a headline price three times DeepSeek V4 Flash's.

Your agent bills while you sleep

This is the one that surprises people. OpenClaw's heartbeat is on by default at 30-minute intervals, and the docs are blunt about it: "Heartbeats run full agent turns. Shorter intervals burn more tokens." That is 48 unprompted model calls a day whether or not you message your agent, plus anything you have on cron.

The same docs give the fix: setting isolatedSession: true drops a heartbeat run from roughly 100,000 tokens to 2,000 to 5,000. On a full-context heartbeat, those 48 daily runs add about 144 million input tokens a month, which is roughly $43 a month on MiniMax M3 and $720 on Claude Opus 5, before you have sent a single message. Check this setting before you blame the model.

Three more traps worth knowing:

Anthropic's cheap cache expires in five minutes. A personal agent sits idle for hours between messages, so the cache is cold nearly every time and you pay a fresh full-price write of the whole history. The one-hour cache exists but writes cost 2x base input. Sparse traffic is the worst case for Anthropic's pricing.

Claude 4.7 and newer use a tokenizer that produces about 30% more tokens for the same text. Every per-token comparison against a Gemini or a DeepSeek understates Anthropic's real cost on those models. It is stated plainly in Anthropic's pricing docs.

OpenAI charges 2x input and 1.5x output above 272K input tokens, for the whole request. An agent whose context quietly drifts past that line doubles its bill with no visible change.

Best OpenClaw model for each job

Prices in this section are each provider's own list price, linked at the point of claim. Agentic scores are Artificial Analysis's agentic benchmark as published through OpenRouter, with their measured cost per request.

Best overall: MiniMax M3

The most-run model on OpenClaw, and it earns it: a 1M context window, image input, tool calling, $0.30 per million input tokens and cache hits at $0.06. On the agentic benchmark it scores 35.4 at $0.50 per request, against Claude Opus 5's 55.3 at $12.61. You give up about a third of the score for a 25x saving, which for daily assistant work is a reasonable trade.

Where it struggles is long unattended chains, where one malformed tool call compounds into five wasted turns. That is a real limitation, and the fix is not a fallback (see below), it is knowing which tasks to hand to a stronger model.

Best when it must not fail: Claude Opus 5

Top of the agentic benchmark at 55.3 and top of the intelligence board. If your agent touches money, production systems or a long unsupervised run, this is what you want holding the wheel.

Note the version. Claude Opus 4.8 is now a legacy model. Anthropic's current line is Fable 5, Opus 5, Sonnet 5 and Haiku 4.5, and OpenClaw's own examples use anthropic/claude-opus-5. Any guide still telling you to run Opus 4.8 in July 2026 has not been checked recently. Setup steps are in our Claude API key guide.

Watch the calendar too: Claude Sonnet 5's $2 and $10 rates are introductory and move to $3 and $15 on 1 September 2026.

Best score per dollar: GPT-5.6 Luna

The quiet winner of this comparison. 45.6 on the agentic benchmark at $0.67 a request, against Opus 5's 55.3 at $12.61. That is 82% of the capability for 5% of the cost. OpenAI lists it at $1.00 in and $6.00 out, and OpenRouter currently routes it for about half that. It is also OpenClaw's declared small-model default for OpenAI, which tells you what OpenClaw's own maintainers think it is good at. Start with our OpenAI API key guide.

Best cheap model for OpenClaw: DeepSeek V4 Flash

$0.14 in, $0.28 out, 1M context, and the second most-used model in real OpenClaw traffic. About $8 a month for the personal-assistant workload above. It has no vision, so pair it with a separate image model if your agent handles screenshots. If your sessions run long, DeepSeek V4 Pro is often cheaper in practice despite the higher sticker price, because its cache hits bill at $0.0036. Setup: OpenClaw with DeepSeek.

Best free model for OpenClaw: Nemotron 3 Super and Ultra

The free models OpenClaw users actually run are NVIDIA's Nemotron 3 free tiers on OpenRouter, nvidia/nemotron-3-super-120b-a12b:free and nvidia/nemotron-3-ultra-550b-a55b:free. Both do tool calling. Ultra's free variant carries a 1M context window.

Read the limits before you commit: OpenRouter caps free models at 20 requests per minute and 50 requests per day until you have bought $10 of credit in your lifetime, after which the daily cap rises to 1,000. Fifty requests a day will not sustain an always-on agent with a half-hourly heartbeat. That is the honest shape of free.

Other genuinely free, tool-capable routes today include openai/gpt-oss-20b:free, google/gemma-4-31b-it:free and inclusionai/ling-3.0-flash:free.

Fastest model for OpenClaw: gpt-oss-120b on Cerebras

522 tokens per second at $0.35 per million input, well clear of the field. Only four of OpenRouter's ten fastest models appear anywhere in OpenClaw's top 20, which tells you agent users do not optimise for throughput. Speed makes a chat reply feel snappy. It does not make a fifteen-step tool chain finish correctly.

OpenRouter's fastest models by throughput, July 2026

Best for long documents: Kimi K3

1M context, cache hits at $0.30, and fourth on the agentic benchmark at 50.1. If your agent reads whole reports or codebases rather than chat messages, the long-context specialists earn their price. We have a Kimi setup walkthrough too.

Best local model for OpenClaw (short version)

A different decision with different constraints, and it comes down to how much RAM you have rather than which model is best in the abstract. Written up separately: best local LLM for OpenClaw, ranked by RAM tier.

OpenClaw model configuration: primary, fallbacks and utility

This is the part almost every model roundup gets wrong. OpenClaw does not ask you for a model. It asks for a primary model, an ordered list of fallbacks, and a separate utility model for cheap internal work like generating session titles and progress narration.

Configuration lives in ~/.openclaw/openclaw.json:

{
  agents: {
    defaults: {
      model: {
        primary: "minimax/MiniMax-M3",
        fallbacks: ["deepseek/deepseek-v4-flash"],
      },
      utilityModel: "openai/gpt-5.6-luna",
    },
  },
}

Or from the terminal:

openclaw models set minimax/MiniMax-M3
openclaw models fallbacks add deepseek/deepseek-v4-flash
openclaw models status

Fallbacks are not a quality ladder

Here is the mistake worth avoiding. Fallbacks fire when the primary is unreachable, not when it gives a bad answer. Putting Claude behind MiniMax as a "fallback for hard tasks" does nothing: Claude only ever runs if MiniMax is down. Fallbacks are a failover path, and OpenClaw rotates auth profiles inside a provider before it moves to the next fallback model at all.

To actually pay up on a specific turn, switch models in the chat itself:

/model list
/model anthropic/claude-opus-5
/model default

A model you select with /model is strict: if it becomes unreachable the reply fails visibly rather than silently degrading to something cheaper. That is what you want when you have deliberately reached for the expensive model. For a permanent split, give a second agent its own model under agents.entries.*.model and route the serious work to it.

The utility model is where quiet spend hides

Leave utilityModel unset and OpenClaw uses the primary provider's small-model default, which on Anthropic means claude-haiku-4-5 at $1 per million input tokens for the privilege of naming your chat threads. Set it explicitly.

Two things that will trip you up

Use rolling aliases so your config does not rot. opus currently resolves to anthropic/claude-opus-5 and follows the family forward, while opus-4.8 stays pinned to that version.

The provider ID is not always the brand name. There is no glm provider in OpenClaw; GLM models live under Z.AI as zai/glm-5.2. Moonshot and Kimi Coding are two separate plugins with two separate keys, and Moonshot must be installed before onboarding.

OpenClaw ships plugins for around 60 providers. A few of the IDs you are most likely to need:

Provider	OpenClaw ID	Example model ref
Anthropic	`anthropic`	`anthropic/claude-opus-5`
OpenAI	`openai`	`openai/gpt-5.6-luna`
MiniMax	`minimax`	`minimax/MiniMax-M3`
DeepSeek	`deepseek`	`deepseek/deepseek-v4-flash`
Z.AI (GLM)	`zai`	`zai/glm-5.2`
Moonshot (Kimi)	`moonshot`	`moonshot/kimi-k2.6`
NVIDIA	`nvidia`	`nvidia/nvidia/nemotron-3-ultra-550b-a55b`
xAI	`xai`	`xai/grok-4.3`
OpenRouter	`openrouter`	`openrouter/auto`
Groq	`groq`	`groq/llama-3.3-70b-versatile`

For the full provider mechanics, custom endpoints and where keys live, that is a separate guide: how to configure an LLM provider in OpenClaw.

How to tell you picked the wrong model

Spend dashboards tell you what a model cost. They do not tell you it is failing. The symptoms of a model that is too weak for an agent harness look like this:

It answers about a tool instead of calling it, describing what it would do rather than doing it.

It emits a tool call as text in the message body, so nothing executes and the reply reads oddly formal.

It calls the same tool repeatedly with the same arguments, ignoring the result it already got back.

Replies get shorter and more confident as the session grows, which usually means context is being compacted away faster than the model can track it.

It starts inventing file paths or command flags that were never in the transcript.

Long chains stall around the same step every time.

Any two of those together mean the model, not the prompt, is the problem. Switch with /model for one session and see whether the behaviour disappears. That is a cheaper experiment than reading another comparison.

What free OpenClaw models actually cost you

Free tiers are the most-searched corner of this topic and the least honestly covered.

Google's free Gemini tier trains on your agent's conversations. Google's own pricing page has a row labelled "Used to improve our products", and for the free tier the answer is Yes on every model. The paid tier says No. The API terms go further: human reviewers may read, annotate and process your input and output, and Google explicitly says not to submit sensitive, confidential or personal information to the unpaid services.

For a chatbot demo that is fine. For an agent wired into your email, your files and your shell, it is a real decision.

If you also see "1,500 requests a day" quoted for the Gemini free tier, Google no longer publishes that number. The current rate limits page says limits depend on your usage tier and can only be seen in AI Studio.

The same scrutiny applies to the cheap models topping the usage table. MiniMax, DeepSeek, StepFun, Tencent and Z.AI are excellent value and their terms are not Anthropic's. If your agent handles anything you would not paste into a public forum, read the provider's data-retention policy before the price, or run the whole thing locally.

The free routes we would actually use: Nemotron 3 Super or Ultra on OpenRouter within the rate limits above, gpt-oss-20b:free, or Z.AI's GLM-4.7-Flash, which is free first-party rather than free with conditions.

How to switch models in OpenClaw

Run openclaw models set provider/model, or open Settings and then Model Providers in the Control UI and use the Default models card. Send a test message to confirm it took. There is a click-by-click walkthrough with screenshots in our guide on how to change the default AI model in OpenClaw.

Because switching costs nothing, start cheap and upgrade only when you can point at a task the model actually failed.

If you would rather not run the server yourself, Agent37's managed OpenClaw hosting is bring-your-own-key on every plan from $3.99 a month, so your key goes straight from your instance to the provider and you pay the provider's rate directly. The $9.99 Plus plan bundles free usage of GPT-OSS 20B, GPT-OSS 120B and DeepSeek V4 Flash, capped per rolling five-hour window and again per week. There is no free tier, so if zero spend is the requirement, the OpenRouter free models above are the honest answer.

Frequently asked questions

What is the best model for OpenClaw?

MiniMax M3 for most people. It is the most-used model in real OpenClaw traffic at 1.08 trillion tokens in 30 days, costs $0.30 per million input tokens, and runs a personal assistant for about $18 a month. Claude Opus 5 is more capable and costs roughly 25 times more per request.

What is the best free model for OpenClaw?

NVIDIA's Nemotron 3 Super and Ultra free tiers on OpenRouter. Both support tool calling and Ultra's free variant has a 1M context window. OpenRouter limits free models to 20 requests per minute and 50 per day until you have purchased $10 of credit, then 1,000 per day.

Is Claude still the best model for OpenClaw?

Claude Opus 5 leads the agentic benchmark at 55.3, and Claude models take more than half of all spend on agent workflow execution across OpenRouter. But every Claude model combined is about 9% of OpenClaw's actual token volume. People pay for Claude on the turns that matter and run something cheaper the rest of the time.

Which model do most OpenClaw users use?

MiniMax M3, by a wide margin, followed by DeepSeek V4 Flash and Step 3.7 Flash. Three quarters of OpenClaw's published token volume goes to models priced under $0.50 per million input tokens.

What is the cheapest model that works with OpenClaw?

DeepSeek V4 Flash at $0.14 in and $0.28 out, with a 1M context window and tool calling. For long-running sessions DeepSeek V4 Pro often costs less in practice, because its cache hits bill at $0.0036 per million.

Which model is fastest for OpenClaw?

gpt-oss-120b on Cerebras at 522 tokens per second. For agent work this matters less than it sounds, since the bottleneck is usually the number of tool calls, not the token rate.

How much does an OpenClaw agent cost per month?

For a personal assistant handling 40 messages a day: about $8 on DeepSeek V4 Flash, $18 on MiniMax M3, $112 on Claude Sonnet 5 and $279 on Claude Opus 5. The default half-hourly heartbeat can add more than the messages do if it runs on full context. Compute is separate and covered in OpenClaw hosting pricing.

Can OpenClaw use more than one model at once?

Yes. OpenClaw takes a primary model, an ordered list of fallbacks and a separate cheap utility model, and individual agents can override the default. Fallbacks fire on provider failure, not on poor answers, so use /model to upgrade a specific session.

What models does OpenClaw support?

Around 60 provider plugins, including Anthropic, OpenAI, Google, xAI, DeepSeek, MiniMax, Moonshot, Z.AI, Qwen, Groq, Cerebras, OpenRouter, Ollama and any OpenAI-compatible endpoint. Model refs take the form provider/model.

Which models support images in OpenClaw?

Among the popular picks: MiniMax M3, Claude Sonnet 5 and Opus 5, the Gemini models, Kimi K3 and Step 3.7 Flash. DeepSeek V4, GLM 5.2 and the gpt-oss models are text only. If your primary is text only, set imageModel separately and OpenClaw routes images to it.

Can OpenClaw use local models?

Yes, through Ollama, LM Studio, vLLM or any OpenAI-compatible endpoint. Expect a real capability gap against hosted frontier models, and pick by available RAM. Details in our local LLM guide.

The bottom line

Run MiniMax M3 as your primary. Set a cheap utility model so you are not paying frontier rates to title chat threads. Check your heartbeat is running isolated, because on full context it can cost more than everything you actually asked the agent to do. Then use /model to reach for Claude Opus 5 on the turns where being wrong is expensive, which is the pattern the spend data shows experienced users already follow.

If money is the hard constraint, DeepSeek V4 Flash will carry a personal assistant for about $8 a month. If it is zero, the Nemotron free tiers work until you hit 50 requests a day.

Then stop optimising and watch your provider dashboard for a week. That is more signal than any leaderboard, including this one.

Last updated 29 July 2026. Usage and throughput figures from OpenRouter's public rankings and its OpenClaw collection, agentic scores from Artificial Analysis via OpenRouter, and prices from each provider's own pricing pages, all captured on that date. Monthly cost estimates use the stated workload assumptions and will move with yours. Model pricing changes quickly, so check before you commit.

Best AI Model for OpenClaw: What 90 Million Real Requests Run