Model Stacks for AI Assistants
A model stack is a curated set of AI models at a set price level that your OpenClaw assistant routes across automatically. Instead of locking you to one model, the stack hands every message to OpenRouter’s auto-router, which works out what kind of task it is (coding, a quick question, research and so on) and sends it to a model in your stack that’s widely relied on for that kind of work. Each stack also sets a price range, so costs stay predictable and match the stack’s intent.
Stacks at a glance
Section titled “Stacks at a glance”Free models only. Explore without spending a credit.
Efficient paid models. Quality stays high, spend stays low.
Balanced quality and speed for everyday work. Includes everything in Hustler.
The best models available. For when quality is everything. Includes everything in Professional.
How we choose models
Section titled “How we choose models”Every stack has different criteria. Before making any changes we cross-reference our own fleet routing telemetry (which models the auto-router actually picks, at what cost), OpenRouter’s category rankings (real-world usage for agent and coding work), each model’s live OpenRouter endpoint specs (pricing, throughput, availability, output-token limits), and independent evaluations from Artificial Analysis.
| Stack | Capability bar | Cost bar |
|---|---|---|
| Wanderer | Solid task completion | Must be :free on OpenRouter |
| Hustler | High success rate, strong value | Lowest cost-per-run with acceptable quality |
| Professional | High success rate, fast throughput | Mid cost, latency matters here |
| Operator | Best available, leads on reasoning | Cost is secondary |
Every candidate is also checked per provider endpoint for output-token limits that would overflow the CX23 VPS — the same model can be safe on one host and unusable on another. See VPS specs and model limits.
Model details and benchmarks
Section titled “Model details and benchmarks”The strongest free model available and the first one Wanderer tries. Answered 10 of 10 calls in our September rate-limit test.
An agentic coding model with image understanding, and Wanderer's image model. Replaced Gemma 4, whose free endpoint was rate-limited on every call.
The fastest free model in our tests (under 200ms). Passed a full agent turn on a real ClawOps bot before joining the stack.
A lighter, quicker Nex model that keeps the bot answering if everything above it is busy.
The cheap, fast floor and the safety fallback for every paid stack. Eight independent endpoints, so it is always reachable.
Cheap, capable and vision-capable. Single Alibaba endpoint with a safe 1M/131K context-output profile.
OpenAI's newest budget tier, half the price of GPT-5.6 Luna and far stronger than the GPT-5 Nano it replaces.
Replaces MiMo 2.5 at the same price. Fastest paid model in our September probes (256ms).
DeepSeek's official replacement for V4 Flash, now with image understanding. Number one in OpenRouter's technology and marketing rankings, and this stack's image model.
Number one in OpenRouter's programming rankings. Fourteen CX23-safe endpoints make it one of the most reliable models in the pool.
Replaces Gemini 3.7 Flash. The fastest model in the stack at around 280 tokens per second, with full audio and video input.
The open-weight flagship behind OpenRouter's "Ox Alpha" stealth launch, served by 18 CX23-safe endpoints.
The best value on the board in September 2026: flagship-level intelligence at under a dollar per million output tokens.
Alibaba's flagship, now pinned to its September snapshot after the original listing was retired.
OpenAI's new mainstream flagship, stronger than GPT-5.6 Terra at a lower price. The ceiling and image model for this stack.
Frontier instruction-following and careful reasoning across coding and agent work. Ten CX23-safe endpoints.
The highest-scoring model on Artificial Analysis in September 2026, and 20% cheaper than the Opus 5 it replaces. Also this stack's image model.
How auto-routing works
Section titled “How auto-routing works”For each message, OpenRouter’s auto-router classifies the kind of task, then picks from the models in your stack that fall inside the stack’s price range, favouring the ones most relied on across OpenRouter for that kind of work. Hustler stays in the economical range, Professional in the middle, and Operator works from the premium models. Wanderer doesn’t use the router: it tries its free models in a fixed order.
The router chooses by the kind of task, not by how hard a particular message is. If you want the most capable models on every message, choose Operator. If you want your credits to go further, choose Hustler.
You don’t configure anything, it happens per message, automatically, within the set you’ve chosen. If the router itself ever has a hiccup, a reliable safety model in the stack catches it, so there’s no downtime.
Credit usage
Section titled “Credit usage”Costs vary by model and message length. Rough per-message estimates:
| Stack | Typical cost per message |
|---|---|
| Wanderer | $0 |
| Hustler | $0.001–$0.005 |
| Professional | $0.003–$0.015 |
| Operator | $0.015–$0.08 |
Check your Dashboard usage tab to monitor actual spend.
Switching stacks
Section titled “Switching stacks”Open your Dashboard, go to Settings, and select a new stack. The change applies immediately, no redeploy needed. You can also ask your bot to use any individual model for a specific conversation without changing your stack.
Model compatibility
Section titled “Model compatibility”Stack models are tested for compatibility with the CX23 VPS before any changes are made. If you use the custom model picker, be aware that some newer 1M context models have a token specification issue that causes immediate overflow errors. See VPS specs and model limits for the full details.