The best model for OpenClaw: what we run, and the default trap

Agent bills are input bills

The economics of running an agent are not the economics of chatting with one. OpenClaw’s own docs describe how context works: the system prompt is rebuilt and sent every run (rules, tool schemas, skills, workspace files), plus the conversation so far, plus every tool result. Output is a few sentences; input is everything. Measured, the spend curve tracks input and cached input almost entirely, so the number that decides your bill is the input price per million tokens, and the cached-input price right behind it.

The default trap

Every setup path OpenClaw offers defaults to a flagship, or to a router free to pick one. Since OpenClaw 2026.9.5, a fresh setup selects openai/gpt-6-astra whether you sign in with ChatGPT or paste an OpenAI API key, per its OpenAI model docs; Astra is OpenAI’s top model, at $10 per million input tokens. Earlier releases picked GPT-5.6 Sol for a subscription sign-in. OpenRouter onboarding defaults to openrouter/auto, a meta-router that picks from dozens of models and bills you at whatever it picked. It can settle on the flagship of the day (in July 2026, one at $5 per million input tokens) for messages that a model a fiftieth of the price could handle, and light use then burns through $10 of credit in about seven days. An update does not move a model you already chose, so this bites on new setups. If you pay per token, the first configuration change worth making is a concrete model choice:

# OpenAI API key, OpenClaw 2026.9.6+
openclaw models set openai/gpt-6-luna
# or in ~/.openclaw/openclaw.json:
# { agents: { defaults: { model: { primary: "openai/gpt-6-luna" } } } }
# OpenRouter
openclaw models set openrouter/openai/gpt-6-luna

GPT-6.1 Sol and Luna in OpenClaw

OpenAI released GPT-6 Sol and Luna on September 22, 2026, and OpenClaw added them the next day in release 2026.9.6, as openai/gpt-6-sol and openai/gpt-6-luna, for an OpenAI API key or a ChatGPT sign-in. The release left the Astra default and existing choices alone, so switching is the command above. On releases before 2026.9.6 the direct ids fail as unknown models, and the OpenRouter line above reaches either one: Luna runs that way as an agent’s model on OpenClaw 2026.9.4.

A week later, on September 29, OpenAI released GPT-6.1 Sol, and its Codex model docs now recommend it for complex coding and agentic work, with Astra kept for the most demanding: “near-Astra performance for complex work at a lower cost than Astra.” In OpenClaw it is openai/gpt-6.1-sol. With an OpenAI API key it works from OpenClaw 2026.9.8, released October 3, 2026; earlier releases have no entry for the id. With a ChatGPT sign-in it takes 2026.9.9, released October 8: that release’s notes add the model to the Codex model list and move the bundled Codex to 0.160.0. 2026.9.8 bundled 0.158.0, and a user report shows ChatGPT refusing GPT-6.1 Sol requests from that version. OpenRouter reaches it on older releases too, because OpenClaw resolves OpenRouter model refs against OpenRouter’s live catalog: openclaw models set openrouter/openai/gpt-6.1-sol. It is not quite a drop-in for GPT-6 Sol. It always reasons: per OpenClaw’s OpenAI model docs, /think off does not turn reasoning off and the lowest effort is low, so a setup that ran GPT-6 Sol without reasoning to keep routine turns cheap loses that option. And it bills cached input at half GPT-6 Sol’s price, with input and output unchanged.

Your account decides which of these you get: openclaw models list --provider openai shows what it offers, and OpenClaw neither grants access a plan lacks nor quietly substitutes another model. On ChatGPT plans, GPT-6.1 Sol is rolling out to Plus, Pro and Business, and to Enterprise and Edu once an administrator enables it; Free and Go are not included at launch. A config that names openai/gpt-6-sol keeps running it: OpenClaw doesn’t move an existing selection, and OpenAI still offers GPT-6 Sol at the same prices, with no retirement date on its deprecations page as of October 2. Before running openclaw models set openai/gpt-6.1-sol, check that the models list shows it as available on your account and release.

What the tiers cost right now

List prices on October 2, 2026, per million tokens (input / output / cached input), from OpenAI’s price list and OpenRouter’s listings:

• GPT-6 Astra (OpenAI flagship): $10 / $50 / $1
• GPT-6.1 Sol (OpenAI’s recommended model for complex work): $2 / $10 / $0.10
• GPT-6 Sol (the Sol it succeeds): $2 / $10 / $0.20
• GPT-6 Luna (OpenAI efficient tier): $0.10 / $0.50 / $0.01
• DeepSeek V4.1 Flash (open-weights efficient model): $0.30 / $1.20 / $0.006 on OpenRouter’s listing, which is DeepSeek’s peak rate; DeepSeek’s own API charges half outside its weekday peak hours
• Claude Opus 5.5 (Anthropic’s recommended default): $4 / $20 / $0.20
• Claude Sonnet 5.5 (Anthropic mid-tier): $2 / $10 / $0.20
• GPT-5-mini (two generations back): $0.25 / $2 / $0.025

Notice the odd one out: GPT-6 Luna, the current generation’s efficient model, is cheaper than an older generation’s small model, and a hundredth of the flagship’s price on the input side, the side agents actually consume. GPT-6 halved Sol’s and Luna’s prices against the GPT-5.6 tiers they replace, so any per-token budget worked out before September 22 is worth redoing. GPT-6.1 Sol then halved Sol’s cached-input price, which counts for more in an agent than in a chat, for the reason below.

The cache decides more than the list price

Most of an agent’s input is the same from one turn to the next: the system prompt, the tool schemas, the workspace files, the conversation so far. As of September 2026, on an OpenClaw 2.x install with its default tools, that repeated part can come to about 28,000 tokens a turn, and prompt caching is what makes it affordable: a short message on the efficient tier costs about a tenth of a cent once the context is cached. The same short message on Gemini 3.5 Flash Lite through OpenRouter, a model in the same price band whose listing carries a cached-input price, costs a cent or more, about ten times as much, because on that route the cache can fail to apply, and the whole context is then billed at full price every turn. A listing tells you the ceiling and the floor, not which one you are paying. After a few short messages, the per-request cost in your provider’s activity log is the fact: if it is not falling once the context repeats, the cache is not applying, and the cheap-looking model is the expensive one. Over a whole job, the number of calls and what each one pays outside the cache can outweigh a twofold gap in list price, as one job measured call by call shows.

The same arithmetic runs while you are away. OpenClaw 2’s heartbeat wakes the agent every thirty minutes by default, whether or not anyone has messaged it, as a full turn on the default model over the same conversation: about 28k tokens a run on a fresh install, and an average of 77k on an agent eleven days into everyday use, where idle heartbeats were half of every token it sent. A default whose cache does not apply pays the full price on every one of those 48 wakes a day; the heartbeat guide has the measured costs and the settings that cut them.

Why the cheapest model is the wrong goal

There is a floor, and OpenClaw’s security docs are the ones who set it: smaller and cheaper models are more susceptible to tool misuse and instruction hijacking under adversarial prompts, and the official recommendation is to use a latest-generation, best-tier model for any bot that can run tools or touch files. The same lesson holds across generations: last-generation small models fumble multi-step setup work that current-generation efficient models handle fine. That’s why the practical sweet spot for a per-token agent is the newest efficient models, not the oldest cheap ones. Our own managed OpenClaw pods start on DeepSeek V4.1 Flash, with the flagship a config line away when a task deserves it. OpenClaw also supports a fallbacks list per agent, so you can pin the efficient model as primary and let the harness fail over rather than silently upgrade.

For the setup and tool-support checks when running a model on your own hardware, see OpenClaw with Ollama.

The subscription exception

All of this math inverts if your ChatGPT subscription powers the agent: a flat-rate plan absorbs the token prices, so the constraint becomes the plan’s usage limits instead. On Plus that is a rolling five-hour window, which OpenAI estimates at about 5–45 Astra messages, 15–160 GPT-6.1 Sol and 350–3,000 Luna; Pro plans currently have no five-hour limit. For scale: the heartbeat alone wakes the agent about ten times in five hours before you send it anything, more than the low end of the Astra range. So on Plus, the default is the model most likely to run the window dry; GPT-6.1 Sol is the everyday choice OpenAI itself recommends (GPT-6 Sol until your OpenClaw lists it), with Astra kept for the hard jobs. One caution: model access follows the plan tier, so if the subscription lapses, the selected model keeps being selected but stops being served, and the agent goes silent with per-request errors until you switch models or renew. The per-token discipline above is for API-key and OpenRouter setups, the ones where every million tokens has a price tag.