Models
One memory. Your choice of model.
Starling MX runs on eight commercial models from Anthropic, Google and OpenAI. Starling recommends one for each task, and you can pick another at any time. Whichever you choose, it reads the same governed memory, uses only what the task needs, and forgets it when the task ends. Smart Enterprise can run with any AI system that connects through MCP, the open protocol a model uses to reach a repository.
Current as of October 2026
The lineup
| Model | Provider | Tier | Use it for |
|---|---|---|---|
| Claude Fable 5 | ProviderAnthropic | TierReasoning | Use it forThe most capable model here, for long, multi-step work. |
| Claude Opus 4.8 | ProviderAnthropic | TierReasoning | Use it forThe newest Opus, for the hardest reasoning. |
| Claude Opus 4.6 | ProviderAnthropic | TierReasoning | Use it forThe previous default, with strong reasoning. |
| Claude Sonnet 5 | ProviderAnthropic | TierReasoning | Use it forThe default: near-Opus quality for everyday work. |
| Claude Sonnet 4.6 | ProviderAnthropic | TierStandard | Use it forThe previous-generation Sonnet, at lower cost. |
| Gemini 3.1 Pro | ProviderGoogle | TierReasoning | Use it forGoogle’s flagship, for very long documents (a context of a million tokens). |
| Gemini 3.6 Flash | ProviderGoogle | TierStandard | Use it forFast, low-cost work. |
| GPT-5.6 Luna | ProviderOpenAI | TierLight | Use it forThe lowest-cost option, by an order of magnitude. |
Reasoning models work a problem through before they answer. Standard models answer directly and cost less. Light is the lowest-cost tier, for quick questions and high volume.
How Starling chooses
- The default is Claude Sonnet 5. Change it for your organization in settings.
- Pick per task. Choose any model from the picker for a single task, or let a session follow your default, so it changes when the default does.
- Starling recommends. For each task, Starling suggests the model that fits. You decide.
- Behind the scenes: Realtime voice runs on OpenAI, which is also the automatic fallback for text generation if a provider is unavailable.
What each provider keeps
None of these providers trains on what Starling sends it, under their commercial and API terms. Anthropic and OpenAI may keep inputs and outputs for up to 30 days for abuse monitoring, and Google for up to 55 days, then delete them.
Anthropic offers zero retention to some enterprise customers, but not for Claude Fable 5 on any plan.
Whatever the model, it sees only the context for one task, never your library. How we protect your data →
Models change. Your memory doesn’t.
Providers release, retire and sometimes suspend models. In June, export controls took two frontier models offline, and work built on them stopped. On Starling, your memory lives in your library, not in the model, so changing models is a choice in the picker, not a migration. We update this page when the lineup changes.
