Model configuration

Model configuration

LLMs are configured at two levels: the Master Orchestrator (MO), which governs the whole bot, and individual AOPs, which can either inherit the MO's model or override it. This page covers both, along with the models being retired.


Watch the tutorial


Master Orchestrator Model

Selects the LLM that powers the Master Orchestrator — the model used to interpret user queries, route them, and generate responses. This configuration is also inherited by every AOP on the bot unless an AOP is given its own model.

Choosing a model: The picker is organised provider-first. Select a provider, then a model. Within each provider, models are listed most capable first, with retiring models at the end. Each model shows a short "best for" line, and newly added models carry a New badge.

ProviderAvailable models
OpenAIGPT 6 Astra (New), GPT 6.1 Sol (New), GPT 5.6 Terra (New), GPT 5.5, GPT 4.1 (retiring)
Azure OpenAIGPT 6 Astra (New), GPT 6.1 Sol (New), GPT 5.6 Terra (New), GPT 5.5, GPT 4.1 (retiring)
AnthropicOpus 5.5 (New), Sonnet 5.5 (New), Sonnet 4.6 (retiring), Opus 4.7 (retiring), Opus 4.6 (retiring), Opus 4.5 (retiring)
Amazon Bedrock (Anthropic)Opus 5.5 (New), Sonnet 5.5 (New), Sonnet 4.6 (retiring), Opus 4.7 (retiring), Opus 4.6 (retiring), Opus 4.5 (retiring)
ModelBest for
GPT 6 AstraFrontier reasoning for the hardest coding and agentic work.
GPT 6.1 SolAstra-class reasoning for everyday production work at a fraction of its cost.
GPT 5.6 TerraBalanced GPT 5.6 tier for everyday production workloads at lower cost.
GPT 5.5High throughput for production-scale workloads.
GPT 4.1Fast and cost-efficient for simple, high-volume tasks.
Opus 5.5Latest Opus: stronger agentic coding and knowledge work, cheaper than Opus 5.
Sonnet 5.5Latest Sonnet: capable everyday agent work at Sonnet pricing.
Sonnet 4.6Balanced speed and quality for everyday tasks.
Opus 4.7High output for production-scale workloads.
Opus 4.6Most capable for the most demanding tasks.
Opus 4.5Strong reasoning for complex tasks at moderate cost.

The default configuration is OpenAI — GPT 5.5. Use the Compare models link in the picker header to view benchmarking data before deciding. See LLM Model Benchmarks.

Model settings:

  • Effort — Controls how thoroughly the model thinks before responding (Low / Medium / High). Higher effort improves reasoning quality on complex requests at the cost of response time. Default: Low. Available on every model except GPT 4.1.
  • Priority processing — Runs requests through a priority lane for faster, more consistent response times. Available on supported models only and Off by default where supported:
ProviderModels with Priority processing
OpenAIGPT 6 Astra, GPT 6.1 Sol, GPT 5.6 Terra, GPT 5.5
Azure OpenAIGPT 5.6 Terra, GPT 5.5
AnthropicOpus 5.5
Amazon Bedrock (Anthropic)None

How it works:

  • Changes apply on save and take effect from the next user request.
  • AOPs inherit this configuration live — changing the model here immediately changes the model every inheriting AOP runs on. AOPs with their own model override are unaffected.
  • Run History and the Debugging Console record the resolved provider and model for every run, so you can always trace which model handled a request.

AOP Model

Selects the LLM this AOP runs on — the model used to follow the AOP's instructions, make decisions, call tools, and generate responses during execution. Every AOP resolves to a model in one of two ways: inheriting from the Master Orchestrator, or holding its own.

Inherited from Master Orchestrator (default)

By default, the AOP uses the Master Orchestrator's model configuration and follows it live — if an admin changes the Master Orchestrator's model, every inheriting AOP switches with it, with no action needed on the AOP. The picker shows this at the top as From orchestrator setting — "Uses the model set in Orchestrator settings. Changes there apply here too." — so you always know which model the AOP will actually run on.

Inheritance is the right choice for most AOPs: it keeps model management in one place and ensures consistent behaviour across the bot.

Or choose a specific model (override)

Override the inherited model when this AOP has different cost, speed, or reasoning needs than the rest of the bot — for example, a simple acknowledgement AOP that doesn't need a heavyweight model, or a complex multi-step decision flow that benefits from deeper reasoning.

The picker is organised provider-first. Select a provider, then a model:

ProviderAvailable models
OpenAIGPT 6 Astra (New), GPT 6.1 Sol (New), GPT 5.6 Terra (New), GPT 5.5, GPT 4.1 (retiring)
Azure OpenAIGPT 6 Astra (New), GPT 6.1 Sol (New), GPT 5.6 Terra (New), GPT 5.5, GPT 4.1 (retiring)
AnthropicOpus 5.5 (New), Sonnet 5.5 (New), Sonnet 4.6 (retiring), Opus 4.7 (retiring), Opus 4.6 (retiring), Opus 4.5 (retiring)
Amazon Bedrock (Anthropic)Opus 5.5 (New), Sonnet 5.5 (New), Sonnet 4.6 (retiring), Opus 4.7 (retiring), Opus 4.6 (retiring), Opus 4.5 (retiring)

An overriding AOP stores its own full configuration (provider, model, and model settings) and is unaffected by Master Orchestrator changes. Re-selecting From orchestrator setting at the top of the picker returns the AOP to live inheritance.

Use the Compare models link in the picker header to view benchmarking data before deciding.

Model settings:

  • Effort — Controls how thoroughly the model thinks before responding (Low / Medium / High). Higher effort improves reasoning quality on complex, multi-step AOPs at the cost of response time. Default: Low.
  • Priority processing — Runs requests through a priority lane for faster, more consistent response times. Available on supported models only (see the table above) and Off by default where supported.

How it works:

  • Changes apply on save and take effect from the AOP's next run — in-flight runs are not affected.
  • The AOP details page shows the effective model, e.g. "GPT 5.5 · OpenAI · High · Priority".
  • Run History and the Debugging Console record the resolved provider and model for every run (marked as inherited where applicable), so you can always trace which model handled a request.
  • Existing AOPs default to inheriting the Master Orchestrator's configuration; nothing changes until you explicitly override.

Retiring models

Models being phased out stay in the picker with an "Available until 30 November 2026" badge. Hover over the badge to see the replacement, e.g. "We're removing this model on 30 November 2026. Opus 5.5 will be selected automatically. Switch earlier if you'd prefer a different model."

Retiring modelProvidersAvailable untilAutomatically moves to
GPT 4.1OpenAI, Azure OpenAI30 November 2026GPT 5.6 Terra
Opus 4.5Anthropic, Amazon Bedrock30 November 2026Opus 5.5
Opus 4.6Anthropic, Amazon Bedrock30 November 2026Opus 5.5
Opus 4.7Anthropic, Amazon Bedrock30 November 2026Opus 5.5
Sonnet 4.6Anthropic, Amazon Bedrock30 November 2026Sonnet 5.5

What happens on the retirement date:

  • Nothing breaks. Any Master Orchestrator or AOP still set to a retiring model automatically runs on its replacement — no action needed.
  • Retired models remain visible in the picker but are greyed out and can't be selected.
  • To choose a different model than the automatic replacement, switch before the retirement date.

Did this page help you?