Supported LLM Providers and Models¶

Our platform integrates with multiple Large Language Model (LLM) providers to offer you flexibility and choice in selecting the most suitable AI models for your needs. Each provider offers unique capabilities and model variations, allowing you to leverage state-of-the-art AI technology through a unified interface. This document outlines the currently supported providers, their available models, and the requirements for using them.

The list above reflects the platform's model catalog at the time of writing; the authoritative list is the model selector inside the app, which is generated from the same catalog.

Models on their way out¶

A model its provider has retired, or announced a shutdown date for, is marked as deprecated in the catalog and stops appearing in the selector. You cannot pick one, but an agent or task configured with it earlier keeps the setting, so a model going away is something to act on rather than wait for.

Two are in that state now:

• GPT-4.1 nano shuts down at OpenAI in October 2026. Anything still pointing at it should move to GPT-5.6 Luna, OpenAI's named replacement and now the platform's default for small utility work. Luna reasons before answering, so expect those calls to take a little longer than nano's did.
• Claude Opus 4.1 is no longer served by Anthropic directly. AWS Bedrock still serves it, which is why it remains in the catalog at all.

Models that are merely older, rather than retired, stay fully available — Claude Fable 5 and Opus 5 have moved behind the 5.1 and 5.5 releases without going anywhere.

Choosing how much a model thinks¶

Many current models can think before they answer, and you can decide how much. When you create or edit an agent and pick a model that supports it, a Reasoning field appears next to Provider and Model. Leave it on Default to let the provider decide, or choose a level. The levels offered depend on the model, from Minimal or Low at the light end up to High, Extra high or Max at the deep end. Some OpenAI models also offer None, which turns thinking off. A model that does not offer a choice simply does not show the field.

Higher levels usually give better answers on hard problems, and they are slower and use more tokens, so they cost more. The field shows which level the provider applies by default when it is known.

Things worth knowing:

• The level belongs to the agent. If you switch the agent to a model that does not offer the level you chose, the setting goes back to Default instead of failing.
• On Claude models, choosing a level turns on the model's extended thinking, and the agent's temperature is then not applied, because Claude does not accept one while thinking.
• Thinking you can see in a chat (the reasoning steps) is separate from this setting, and showing it depends on your role's Show AI Reasoning Details permission.

How it works¶

When using these models in your application:

1. Each model requires its corresponding credential setting to be set in your user or group settings
2. The model name must be prefixed with the provider's prefix, separated by a colon (provider:model). Examples:
   • Google: google_genai:gemini-3.5-flash
   • Anthropic: anthropic:claude-sonnet-5
   • AWS Bedrock: bedrock:claude-sonnet-5
   • OpenAI: openai:gpt-5.5
3. The system will automatically:
   • Validate the presence of the required credential setting
   • Strip the provider prefix when needed (for providers like Anthropic and Groq)
   • Initialize the appropriate client with the correct endpoints and configurations

AWS Bedrock Setup¶

AWS Bedrock models run through your AWS account rather than a direct provider API key. The provider carries the Claude families plus GPT-5.6 Sol/Terra/Luna, Mistral Large, DeepSeek V3.2, and Kimi K2.5. Before selecting one:

1. Enable access to the required model in your AWS Bedrock account.
2. Create AWS credentials with permission to invoke that model.
3. Add AWS_BEDROCK_CREDENTIALS as a User Variable or Group Variable using this exact pipe-delimited format:

region|access_key_id|secret_access_key[|session_token]

For example:

us-east-1|AKIAEXAMPLE|secret-access-key

The optional fourth field is required for temporary credentials that use an AWS session token.

The Claude and GPT-5.6 models are invoked through cross-region inference profiles, so their region must be in a supported inference geography: US/Canada, Europe, Japan, or Australia/New Zealand. Mistral Large, DeepSeek V3.2, and Kimi K2.5 are invoked on demand with the plain Bedrock model id, so they simply need to be enabled in the region you configure.

Bedrock Claude supports vision, reasoning, and native PDF input. It does not provide Anthropic's hosted MCP connector or provider-side web search, web fetch, and code-execution tools. PrimeThink can still use separately configured external web-search tools and its sandbox capabilities.

The non-Claude Bedrock models go through Amazon's Converse API. They also have no provider-side builtin tools, and PDFs are handled by extracting their text rather than being passed to the model as a native document. GPT-5.6 on Bedrock is a reasoning model with vision; Mistral Large accepts text only.

OpenRouter Setup¶

OpenRouter is an aggregator: one API key reaches models from several vendors through a single endpoint. Add OPENROUTER_API_KEY as a User Variable or Group Variable, then pick a model whose name keeps the upstream vendor in it, for example openrouter:anthropic/claude-sonnet-5 or openrouter:google/gemini-3.6-flash. Because the vendor is part of the model name, searching the model selector for "anthropic" shows both the direct Anthropic models and the routed ones.

A routed model is not identical to the same model bought directly:

• No hosted MCP. MCP capabilities are skipped on OpenRouter models. See MCP Capabilities.
• Web search can use OpenRouter's own search plugin, or the external search providers you configure such as Tavily, Perplexity, or Serper, which take precedence. See Web search on OpenRouter models for the cost implications.
• PDF handling depends on the route. Where the route reads PDFs natively, the file is sent as a document; otherwise PrimeThink extracts the text first, deliberately, because OpenRouter would otherwise bill its own paid OCR engine.
• Context and output limits follow the route that OpenRouter actually serves, which can be smaller than the model's headline figures.
• The key is required explicitly. A missing user or group setting is an error rather than a silent fall-back to a platform-wide key, so usage is never billed to the wrong account.

Future Updates¶

We are actively working on expanding our supported providers and models. Future updates will include: - Additional language model providers - New model versions as they become available - Enhanced capabilities and specialized models - Support for more regional endpoints and deployment options

Please check our documentation regularly for updates on newly supported models and providers.