Today we're shipping support for 50+ models across 8 providers — every flagship LLM you'd actually want to run an agent on, one CLI command (or one dropdown click) away.

Anthropic, OpenAI, Google, xAI, DeepSeek, Groq, Alibaba, ZhipuAI. Pick any of them per agent, switch whenever something better ships, and the rest of your Lua setup — Skills, Tools, Webhooks, Spaces — keeps working exactly as it did.

The providers we support

The lineup, with what we'd reach for each one:

  • Anthropic — Claude Opus 4.7 (the new flagship, 1M context), Opus 4.6 / 4.5 / 4.1, Sonnet 4.6 / 4.5, and Haiku 4.5. Our default pick when an agent has to plan, reason, or hold a long conversation.
  • OpenAI — GPT-5, GPT-5.4, and their Mini / Nano / Pro variants. Plus GPT-4.1 (1M context) and the o3 / o4 reasoning family.
  • Google — Gemini 2.5 Flash, 2.5 Pro, 2.5 Flash Lite, and the 3.x preview family. google/gemini-2.5-flash is the platform default — it hits the best price/performance balance for most agents.
  • xAI — Grok 4 (256K context), Grok 4 Fast (2M context for long-document retrieval), Grok 4.20 Multi-Agent (2M context, purpose-built for multi-agent workflows), plus Grok 3 / 3 Mini / 3 Mini Fast.
  • DeepSeek — deepseek-chat and deepseek-reasoner.
  • Groq — Llama 3.1 Instant, Llama 3.3 Versatile, Llama 4 Scout, Compound, Compound Mini, GPT-OSS-120B and GPT-OSS-20B. Worth using when you need raw speed (the 20B model runs at ~1000 t/s).
  • Alibaba — Qwen 3 Max, Qwen 3.6 Plus, Qwen 3.5 Flash (1M context with thinking mode), Qwen 3 235B, Qwen 3 32B, QwQ Plus.
  • ZhipuAI — GLM 5.1, GLM 5, GLM 5 Turbo, GLM 4.7 / 4.7 Flash / 4.7 FlashX, GLM 4.6, GLM 4.5 / 4.5 Air.

Set it from the CLI

The new lua models command lists what's approved, lets you pick, and writes the choice into your agent config:

lua models                                    # list approved models, grouped by provider
lua models set                                # interactive picker
lua models set --model anthropic/claude-opus-4-7
lua models unset                              # revert to the platform default
lua models list --json                        # structured output for scripting
lua models CLI output showing approved models grouped by provider
lua models — the approved list, grouped by provider, with the current model highlighted.

Setting a model writes model: 'provider/model-id' into your agent's LuaAgent config and syncs it to the server. Unsetting removes it from both. The server is authoritative, so anyone on your team running lua models sees the same currently-active model.

Set it from the dashboard

Same list, same defaults, but for teams that prefer to point and click — every agent in the admin dashboard now has a model dropdown right alongside the persona settings.

Lua admin dashboard model picker showing provider and model dropdowns
Pick a provider, pick a model, save. Same approved list as the CLI.

How we picked the list

We don't allow every model under the sun, and that's deliberate. We hand-pick the latest flagship models from each provider, drop deprecated ones as they age out, and add new ones the day they ship.

The reasoning is simple: you shouldn't have to research which OpenAI model is current or whether a particular Gemini variant has been quietly retired — if it's in lua models, it's a model we'd actually deploy on a production agent today.

If you don't see a model you need, tell us. We add models on request.

Direct to the provider

Wherever possible, your agent's requests go straight to the model provider — Anthropic, OpenAI, Google, xAI, and the rest. Going direct rather than through an aggregator gives you the lowest latency, the best rate limits, and a cleaner path overall. It's a small detail that adds up across millions of agent turns.

About training

Across every provider integration, we opt out of letting the provider use your users' conversations to train their models. Whatever your customers say to your agent stays with you and the model serving the request — it doesn't get folded into someone's next-generation training run.

What this unlocks

A few patterns we've seen work well:

  • High-volume support agent — Claude Haiku 4.5 or Gemini 2.5 Flash. Cheap, fast, plenty smart for triage.
  • Reasoning-heavy planner — Claude Opus 4.7 or GPT-5.4 Pro. When the agent has to multi-step plan, write code, or do careful analysis.
  • Long-document retrieval — Grok 4 Fast (2M context) or Gemini 2.5 Pro (1M). Drop in a whole product manual without chunking.
  • Multi-agent orchestration — Grok 4.20 Multi-Agent (2M context, purpose-built for the job) or Groq Compound for fast tool-using sub-agents.
  • Voice agents — Gemini 2.5 Flash for the conversational core, with the rest of the Voice stack handling LiveKit / Recall transport.
  • Reasoning on a budget — DeepSeek Reasoner or Qwen 3 Max. Strong chain-of-thought without flagship pricing.

You can run different agents on different models, switch a single agent between models in seconds, and A/B-test by deploying two versions side by side.

Get started

Update the CLI to v3.10.0 or later, then:

lua models                  # see what's approved
lua models set              # pick one for your agent

Or open any agent in the Lua admin dashboard and use the model dropdown.

The full provider/model matrix lives in the model selection docs, and the v3.10.0 changelog is here.

Pick the model that fits your agent. Switch when something better ships. Don't see one you need? Tell us — we add models on request.