Models
Connect and switch how PandaOS is powered: PandaOS Cloud, your own API keys, local models, or a provider subscription.
Ways to connect
PandaOS supports six model connections. You can set up all of them and switch between them at any time. Each connection has its own model catalog, and the models available in your chat picker reflect whichever connection is currently active.
| Connection | How it works | Best for |
|---|---|---|
| PandaOS Cloud | Included with your subscription. Active by default after onboarding. You spend Credits as you work. | Getting started immediately with no configuration. |
| Your API Keys | Add any number of providers (OpenAI, Anthropic, Google, Azure, OpenRouter, Mistral, and more) with your own credentials. Each provider may need an API key, base URL, region, or service-account JSON. You register which models each provider should expose. | Full control over providers, pricing, and model selection. |
| Local | PandaOS scans for Ollama, LM Studio, KoboldCpp, and llama.cpp on their default ports and connects to whichever ones are running. You can add your own endpoint for any server that speaks /v1/chat/completions. | Keeping prompts on your machine, or working offline. |
| OpenCode Go | A subscription from opencode.ai, served through an OpenAI-compatible endpoint. Works with the OpenCode harness and with Claude Code. | A single subscription across many models. |
| OpenAI Codex | Sign in with your ChatGPT account or paste an OpenAI API key (sk-...). Billed through your OpenAI account. | Teams already invested in the OpenAI ecosystem. |
| Claude | Run on your own Claude subscription, or on Claude models hosted in AWS Bedrock using a bearer token and region. | Teams on an existing Anthropic or AWS contract. |
Inside a connection
Every connection opens on three tabs, in the order you normally work through them:
| Tab | What you do there |
|---|---|
| Account | Sign in, paste a key, or add a region. A connection can hold several accounts, so you can keep a personal and a work login side by side and switch without signing out. |
| Models | Choose which of the connection's models are shown in your pickers. Hiding the ones you never use keeps the composer short. |
| Tiers | Say which model each tier runs. Tiers are what agents, automations and new chats resolve against. |
Selecting a connection makes it active immediately. Your next message uses it, with no restart.
Model tiers
Rather than one "default model", PandaOS resolves work through four tiers, set per connection. A connection you are not currently running on still has live tiers, because an agent or automation bound elsewhere resolves against the engine it runs on.
| Tier | What runs on it |
|---|---|
| High | Deep work: planning, review, hard problems. Used by agents bound to the High tier. |
| Medium | Everyday work, and the default for new chats. Changeable per chat in the composer. |
| Low | Quick, simple tasks. |
| Utility | Chat titles, commit messages, knowledge extraction. Nothing selects it, it just runs. |
On PandaOS Cloud the tiers arrive pre-filled from the server's routing configuration and are re-seeded when that changes, but any tier you set by hand is respected from then on.
Each connection also carries its own reasoning effort default (Low through Max). Higher effort produces more thorough responses and uses more tokens. You can override effort per message in the composer, and an agent can carry its own effort override. See Agents for binding an agent to a tier or to one locked engine and model.
Auto-routing
With Auto selected in the composer, PandaOS picks a model per message based on how complex the request looks, instead of you choosing every time. Which model each complexity level maps to is exactly the tier configuration above, so tuning Auto means tuning your tiers.
The model catalog
Which models you can pick depends on the connection that is active. PandaOS Cloud serves a fixed catalog we host and pay for. Every other connection serves whatever your own account or machine exposes, so those catalogs are as wide as your keys are.
Per-token prices are not listed here. On PandaOS Cloud you spend a share of your plan's allowance rather than paying per model, and on every other connection you are billed by the provider directly, on their rates.
PandaOS Cloud
Fourteen chat models, several of them served from two providers at once so a provider outage does not take the model down with it.
| Model | Served by | Notes |
|---|---|---|
| GPT-5.5 | OpenAI | Auto's medium and complex route on OpenAI |
| GPT-5.4 | OpenAI | Previous GPT generation |
| GPT-5 mini | OpenAI | Auto's simple route on OpenAI |
| GPT-5.3 Codex | OpenAI | Tuned for code |
| DeepSeek V4 Flash | Fireworks, Together | Auto's default and coding route |
| DeepSeek V4 Pro | Fireworks | |
| MiniMax M2.7 | Fireworks, Together | Auto's simple and medium route on both pools |
| MiniMax M3 | Fireworks, Together | Auto's vision route, for images and visual work |
| GLM 5.2 | Fireworks | Auto's complex route on Fireworks |
| GLM 5.1 | Together | |
| Kimi K2.6 | Fireworks, Together | |
| Kimi K2.7 | Fireworks, Together | Auto's complex route on Together |
| Kimi K3 | Fireworks, Together | Frontier-priced, so it draws down faster |
| Qwen3.7 Plus | Fireworks |
Claude models are not part of the PandaOS Cloud catalog. To run Claude, connect your own Claude subscription, an Anthropic key under Your API Keys, or Bedrock.
Voice, embeddings and image generation run on their own models, covered below.
Claude
The same eight models whether you arrive by Claude subscription, an Anthropic key, or Bedrock. Reasoning effort is per model, not universal, so the ladder below is what the composer will actually offer you.
| Model | Reasoning effort | Notes |
|---|---|---|
| Claude Fable 5 | Low, Medium, High, Max | Latest Fable generation |
| Claude Opus 5 | Low, Medium, High, XHigh, Max | Newest Opus, most capable |
| Claude Opus 4.8 | Low, Medium, High, XHigh, Max | Hidden by default while Opus 5 is present |
| Claude Opus 4.7 | Low, Medium, High, XHigh, Max | Hidden by default while Opus 5 is present |
| Claude Opus 4.6 | Low, Medium, High, Max | Hidden by default while Opus 5 is present |
| Claude Sonnet 5 | Low, Medium, High, Max | Near-Opus intelligence |
| Claude Sonnet 4.6 | Low, Medium, High, Max | Hidden by default while Sonnet 5 is present |
| Claude Haiku 4.5 | Not applicable | Fastest and cheapest, no effort levels |
Superseded generations are hidden from the composer until you opt back in from the connection's Models tab. Hiding is conditional: a model is only hidden while its successor is in the same list, so a connection whose newest Sonnet is 4.6 still shows it.
OpenAI Codex
Codex ships no fixed list. PandaOS reads the catalog from the Codex CLI on the account you signed in with, and remembers the last good answer per account, so the models you see are the ones your ChatGPT plan actually entitles you to.
The current generation is GPT-5.6, in Sol, Terra and Luna variants. Earlier ones (GPT-5.5, 5.4, 5.4 mini, 5.2, and 5.3 Codex Spark) are hidden by default once a newer generation appears in the same list.
Codex resolves reasoning effort against its own ladder rather than the shared one, and a model that does not accept an effort level simply will not be sent one.
Your API keys, Local, and OpenCode Go
These three have no catalog of ours to list, because the catalog is yours.
| Connection | Where the list comes from |
|---|---|
| Your API Keys | The providers you add and the models you register on each. Because one list can mix keys, the picker groups models under the provider paying for them. |
| Local | Whatever is loaded in Ollama, LM Studio, KoboldCpp, llama.cpp, or your own endpoint. Grouped by runner, since the same model name can exist on two of them. |
| OpenCode Go | The models your opencode.ai subscription serves. |
Voice, embeddings and image generation
The Models page also configures the models that are not chat models. Each is set independently of your coding engine, so switching harness never disturbs them.
| Capability | What it powers | Providers |
|---|---|---|
| Voice | Dictation in the composer and on mobile. | PandaOS Cloud, or your own Together.ai or OpenAI key. |
| Embeddings | Semantic search across your sessions and project knowledge. | PandaOS Cloud, or your own key. |
| Image generation | The text-to-image tool in chat, generated pictures render inline in the thread. | PandaOS Cloud, or your own OpenAI or Together key. In Codex with a ChatGPT login, PandaOS uses Codex's own hosted image tool and it bills to your ChatGPT plan. |
Image generation is off until you turn it on from the Tools menu in the chat bar. The first PandaOS Cloud generation asks you to acknowledge that it spends Credits.
Provider preference
Several PandaOS Cloud models are served from two providers at once, for example DeepSeek V4 Flash and Kimi K2.7 run on both Fireworks and Together. Where that happens you can set a preference order, and PandaOS routes to the first available provider in it. That gives you a say in latency and in which pool absorbs the load, without changing model.
The same applies when a model you have on PandaOS Cloud is also reachable through a key of your own.
Related
- Credits - your PandaOS Cloud balance and top-ups.