Models
Best & Cheapest Models for OpenClaw — Cost vs Quality Guide (2026)
OpenClaw lets you choose any AI model as your main "brain" and a separate, cheaper model for low-stakes background work. This guide breaks down which models give the best quality, which are cheapest, and how to combine them to keep costs low without sacrificing answers that matter.
How OpenClaw Uses Models
OpenClaw separates model usage into two tiers:
- Primary (main) model — handles user conversations, tool calls, and complex agentic reasoning. This is where quality matters most.
- Background / heartbeat model — handles periodic pings, lightweight summaries, memory indexing, and other low-stakes internal tasks that run even when you are not actively chatting. This runs far more often, so its per-token cost compounds quickly.
Because of this split, your real monthly cost is not just the main model's price per token — it is the main model cost plus the background model cost multiplied by how often your agent runs background tasks. Choosing a cheap heartbeat model can cut total spend by 40–70% even if you keep an expensive main model.
You set these independently in OpenClaw's model settings. The main model and the heartbeat model can come from different providers. OpenRouter makes this easy: one API key gives you access to dozens of models across providers, so you can mix DeepSeek for background tasks with Claude Sonnet for chat, all from a single key.
Model Cost vs Quality — At a Glance
The table below shows popular models you can use with OpenClaw, their rough cost tier, and the scenarios they excel at. Cost tiers are relative (not exact prices — those change frequently) and assume access via OpenRouter or the provider's direct API.
| Model | Cost Tier | Best For |
|---|---|---|
| DeepSeek V3 | Very low | Strong general reasoning, coding, long context. Excellent heartbeat or primary model for cost-focused setups. |
| Qwen (2.5 / 3 series) | Low | Multilingual tasks, coding, tool use. Great cheap background model. Direct API or via OpenRouter. |
| Gemini 2.x Flash | Very low | Fast, cheap, good at structured output and summarization. Ideal heartbeat / background model. |
| Gemini 2.x Pro | Medium | Long context, multimodal reasoning, solid agentic tasks. Good primary for mixed workloads. |
| Claude Sonnet (latest) | Medium | Outstanding agentic reasoning, tool use, and code. Top pick for a primary model when quality matters. |
| Claude Opus (latest) | High | Best-in-class reasoning and writing quality. Use when accuracy is non-negotiable; too expensive for heartbeat. |
| GPT mid-tier (e.g. gpt-4.1-mini) | Medium | Reliable general assistant, good tool use, familiar API. Solid primary if you already hold an OpenAI key. |
| ChatGPT (via your plan) | Low (subscription) | Excellent value if you already pay for ChatGPT Plus. OpenClaw can route chat through Codex OAuth — no extra per-token cost on your plan. |
Cheapest Setups
Route via OpenRouter for One-Key Access to Cheap Models
OpenRouter lets you access DeepSeek, Qwen, Gemini Flash, and many others with a single API key. You only pay for what you use — no subscriptions per provider. In OpenClaw, set your provider to OpenRouter and pick any cheap model for the heartbeat slot.
For example: set deepseek/deepseek-chat or google/gemini-flash-2.0 as your background model in OpenClaw's model settings. Both are available on OpenRouter at very low per-token rates.
Use DeepSeek or Qwen Directly for Maximum Savings
If you want the absolute lowest cost on background tasks, add a DeepSeek or Qwen API key directly (not via OpenRouter). Their direct pricing is among the cheapest available. DeepSeek V3 in particular punches well above its price — many users find it matches or exceeds GPT-4-class quality on coding and reasoning tasks while costing a fraction of the price.
Set a Cheap Model as the Heartbeat / Background Model
In OpenClaw's model configuration, look for the heartbeat or background model setting (sometimes labelled as the "secondary" or "utility" model depending on your version). Point this at Gemini Flash, DeepSeek V3, or Qwen while keeping your main model at whatever quality level you prefer. This single change typically cuts total token spend significantly because background pings accumulate through the day even when you are not chatting.
Best-Quality Picks for Hard Agentic Tasks
If your OpenClaw agent is doing heavy work — multi-step research, writing long documents, running complex tool chains, or operating autonomously for hours — quality on the primary model matters. In that case:
- Claude Sonnet is the go-to for agentic tasks. It follows tool-use instructions reliably, handles long multi-turn context well, and is noticeably better than most alternatives at not getting stuck in reasoning loops. It sits at a medium price tier that most users find acceptable.
- Claude Opus is the ceiling for quality. If you need the best possible output — for example, production code review, nuanced writing, or high-stakes decisions — Opus delivers. Its cost is high, so most users reserve it for their primary chat model only and never use it for heartbeat tasks.
- Gemini 2.x Pro is a strong alternative if you are Google-ecosystem-heavy or want very long context windows at a medium price.
For users who already have ChatGPT Plus: OpenClaw supports ChatGPT via Codex OAuth, meaning your OpenClaw chat can run through your existing subscription at no additional per-token cost. This is often the best-value primary model option if you already pay for the plan — there is no reason to move away from it.
Free Tier Note
Several providers offer free tiers or trial credits that work with OpenClaw:
- Google AI Studio provides a free tier for Gemini models with generous rate limits for personal use.
- OpenRouter gives new accounts some free credits to test models before committing.
- DeepSeek and Qwen offer free tiers with rate-limited access on their direct APIs.
Free tiers are great for testing your OpenClaw setup. For sustained 24/7 agent operation, paid access is more reliable — rate limits on free tiers can interrupt your agent during peak hours.
Recommended Combinations
- Budget setup: DeepSeek V3 as primary + Gemini Flash as heartbeat via OpenRouter. Very low total cost, strong quality for most tasks.
- Balanced setup: Claude Sonnet as primary + DeepSeek V3 as heartbeat via OpenRouter. Excellent agentic quality, low background cost.
- Quality-first setup: Claude Opus as primary + Gemini Flash as heartbeat. Best output quality with cost contained on the background layer.
- Existing ChatGPT Plus subscriber: ChatGPT (Codex OAuth) as primary + Gemini Flash or DeepSeek V3 as heartbeat. No extra cost on the primary; cheap background.