Skip to main content

​
Overview

Aster Agents supports bring-your-own-API-key across multiple providers. You connect your own API keys in Control Hub > Providers, then select a model when building each agent. This gives you full control over cost, performance, and provider choice.
Don’t use outdated models. Models like GPT-4o or GPT-4.1 are previous-generation and significantly underperform current models on agentic tasks. Always use the latest models listed below.

​
Supported Providers

Add your API key in Control Hub > Providers to unlock that provider’s models across all your agents.

​
Preferred Provider

In Control Hub → Models, an admin picks the organization’s preferred provider: Anthropic, OpenAI, xAI, Google, or an Azure OpenAI deployment. Aster Agents keeps a current model at three price tiers on each provider: When a provider releases a newer model, these move to it automatically. The Models page shows which model each tier uses right now. Agents use the same tiers. In an agent’s model menu, pick a tier under Always current and the agent moves to newer models along with your organization, or pick one under Specific model to keep the agent on exactly that model. New agents start on Standard. If a newer model doesn’t support an agent’s saved reasoning effort, the agent runs at that model’s default instead. If the preferred provider isn’t connected, Aster Agents uses the first connected one in this order: Anthropic, OpenAI, xAI, Google. Azure deployments are specific to your account, so an Azure organization picks the deployment that serves every tier.

​
Model Identifiers

When configuring a model — in the UI, the API, or via the manage_agents tool — the platform expects the provider:model-id format, not the display name. "claude-sonnet-4-6" will not work; "anthropic:claude-sonnet-4-6" will. The model picker in each agent’s settings lists every model your connected providers offer. Three models cover almost every agent. Start at the top and move only when you have a reason.

Start with Claude Opus 5.5

Frontier-grade judgment at 4/4 / 20 per 1M tokens with a 1M-token context window. Anthropic rates it at Claude Fable 5.1’s level on most work.

Choose GPT-6.1 Sol if you're on OpenAI

The same tier for half the price — 2/2 / 10 — and the cheapest frontier model on the platform.

Upgrade to Claude Fable 5.1 for the hardest work

Reach for it on long-horizon agentic runs and reasoning where Opus 5.5 at high effort still falls short. It costs 2.5x Opus 5.5, so make it a per-agent decision, not a default.

Claude Opus 5.5

Best all-around. The default choice for most agents. Near-flagship intelligence at mid-tier pricing, with cache reads at 5% of the input price.

GPT-6.1 Sol

Cheapest frontier. OpenAI’s current workhorse — roughly half the error rate of GPT-5.6 Sol at half the price.

Claude Fable 5.1

Maximum capability. For deep reasoning, long-horizon planning, and tasks where subtle errors have real consequences.
Best for: General-purpose agents, CRM workflows, document analysis, data extraction, customer-facing agents, scheduled tasks, multi-tool orchestration.Why Opus 5.5 over Fable 5.1? Opus 5.5 lands at Fable’s level on most work at 40% of the price. For the bread-and-butter of agent tasks — querying databases, filling out forms, searching knowledge bases, writing emails — the difference is imperceptible.Cheap context reuse. Cache reads cost 5% of the input price rather than the usual 10%, so agents with large system prompts or knowledge base content get cheaper the longer a thread runs.
Best for: Coding and code review, research agents, teams already invested in the OpenAI ecosystem, and anywhere frontier quality has to fit a tight budget.Why GPT-6.1 Sol? It costs the same as GPT-6 Sol with cheaper cache reads, and has half the list price and roughly half the mistakes of GPT-5.6 Sol. There is no reason to start a new agent on GPT-5.6.
Best for: Legal document analysis, financial modeling, deep research agents, complex multi-agent workflows, agents that have to handle edge cases gracefully.When to choose Fable: When your agent regularly encounters situations that require judgment calls — interpreting vague requirements, reconciling conflicting information, or producing output where subtle errors have real consequences — and Opus 5.5 at high effort isn’t getting there.
Provider tools aren’t tied to a specific model. Web Search, Code Interpreter, and Image Generation are available on every current OpenAI model, and the Google and xAI tools on any model from those providers. (The legacy GPT-4o and GPT-4.1 models are the exception — they run on an older API that can’t carry server-side tools.) Picking GPT-6.1 Sol over an older GPT-5.x model costs you nothing in tool access — see Provider Tools below.

​
Budget-Friendly Options

For high-volume or cost-sensitive workloads:
These models are excellent for multi-agent workflows where a cheaper model handles routine subtasks (classification, data formatting, simple lookups) while a flagship model handles the complex reasoning. GPT-6 Luna is both cheaper and more capable than the older GPT-5.4 Mini and Nano — there’s no reason to start a new agent on those.

​
Choosing by Use Case

​
Reasoning Effort

Reasoning models think before they answer. Reasoning effort sets how much, per agent. It’s in the agent’s settings under the model picker, and it’s the reasoningEffort field on manage_agents and the agents API. Reasoning is billed as output tokens, so effort moves cost directly. In our tests, High used 2–5x the reasoning of Low on the same question. The levels offered depend on the model, and models without the setting don’t show it (for example Claude Haiku 4.5 and Sonnet 4.6, and Grok or Gemini deployed through Azure). Gemini 3.1 Pro offers Low and High only. In the agent form, switching to a model that doesn’t offer the saved level resets it to Model default. Through the API or manage_agents, the mismatch is rejected until you change or clear the level.

​
Cost Optimization Tips

  1. Start with Opus 5.5 or GPT-6.1 Sol — they handle 90%+ of agent tasks well. Only move up to Fable 5.1 if you see quality issues on specific tasks.
  2. Use cheaper models for subagents — in multi-agent workflows, the orchestrator can run on a flagship model while worker agents run on GPT-6 Luna or Haiku 5.5.
  3. Prompt caching saves money — both providers automatically cache repeated context, and cached input costs 90% less than fresh input on most current models (95% less on Claude Opus 5.5 and GPT-6.1 Sol, 97.5% less on Claude Fable 5.1; less on Grok and Fireworks models). Agents with large system prompts or knowledge base content benefit significantly.
  4. Turn reasoning effort down on routine agents — a search, triage, or routing agent rarely needs High. See Reasoning Effort.
  5. Keep system prompts focused — longer prompts cost more on every message. A concise, well-structured prompt outperforms a verbose one and costs less.

​
Provider Tools

Some providers include built-in tools that extend your agent’s capabilities beyond Aster’s standard tool library: These tools appear automatically when you select a model from the corresponding provider. They can be enabled alongside Aster’s standard tools.