Which LLM Does Agentforce Use?

By default, Agentforce uses a Salesforce-managed mix of models that currently includes OpenAI GPT-4o. In Setup, an admin can switch the whole org to Anthropic Claude on Amazon Bedrock or Google Gemini 3.5 Flash on Vertex AI, and can override that choice for a single agent or subagent. Koa, Salesforce's own CRM reasoning model, is in pilot.

That is the whole answer for most buyers. The detail matters because Salesforce's documents were written at different times. Trailhead describes GPT-4o as the default model for reasoning-engine calls. The developer documentation, which is the most current, says the default is a "managed mix of trusted models (currently including GPT-4o)". The Agentforce Privacy FAQ, published June 2025, says customers who want to know which third-party LLM their implementation uses should contact their Account Executive. This page follows the developer documentation.

Version note on the AWS-Hosted option: Salesforce's developer documentation names Claude Sonnet 4, which Salesforce reroutes to Claude Sonnet 4.6. Salesforce Help's Setup article names Claude Haiku 4.5. The two documents were updated at different times, so confirm the current model in your org's Setup before you rely on it.

The org-level option drives the reasoning engine and agent actions. Custom actions that run prompt templates, Apex, or the Models API can use any Salesforce-managed or bring-your-own model regardless of that setting.

For the product explainer, see What is Agentforce?. For the glossary entry, see LLM.

Salesforce publishes one Supported Models table for Agentforce, Prompt Builder, and the Models API. As of 17 September 2026 it lists these families. Beta models have lower rate limits and Salesforce recommends enabling them in sandbox or development orgs only.

Older models are rerouted rather than removed: Claude 3 Haiku to Haiku 4.5, Claude Sonnet 4 to Sonnet 4.6, GPT-4 and GPT-4 Turbo to GPT-4o, Gemini 2.0 Flash to 2.5 Flash. Salesforce's August 2026 release notes state that Gemini 2.5 requests reroute to Gemini 3.5 on 20 October 2026. Salesforce's Agent Script documentation recommends GPT-4.1, Claude Haiku 4.5, or Gemini 3.5 Flash for agents because those have been tested with agents.

Yes. Salesforce's Agent Script documentation states that Agentforce uses the org-level model for all agents and subagents by default, and that model_config overrides it for a specific agent or subagent. The Agentforce Builder exposes the same agent- and subagent-level picker.

• A subagent-specific model takes precedence over an agent-specific model.
• An agent-specific model takes precedence over the org setting.
• Any supported model API name can be specified. The Supported Models page also lists the API name of any custom model configuration in AI Models.
• Some templates, including the Service agent, default to Salesforce's EinsteinHyperClassifier model for subagent routing. You can keep it or specify a different model in the agent router.

Salesforce's own example runs the org on Salesforce Default, one agent on Claude Haiku 4.5, and one of its subagents on Gemini 3.1 Pro. Mixing providers inside one agent is supported. The glossary entry on model routing covers the general pattern.

Koa is Salesforce's first CRM reasoning model for Agentforce, announced on 15 September 2026 during Dreamforce week. Salesforce built it by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic CRM dataset. Salesforce controls the weights and runs post-training and inference inside its own trust boundary, and states that no customer data is used in training. On Salesforce's own CRM benchmark it reports three times fewer errors than the models it compared against.

Until Salesforce says otherwise, treat Koa as a fourth option alongside the managed mix, Claude on Bedrock, and Gemini, not as a replacement for any of them. The base Nemotron 3 Super 120B and Nano 30B models are separately listed as beta models on Amazon Bedrock and are not Koa.

Partly. Salesforce's bring-your-own-LLM (BYOLLM) feature lets you add a model from Amazon Bedrock, Azure OpenAI, OpenAI, or Google Vertex AI using your own credentials, and the LLM Open Connector lets you connect other endpoints, including custom-built models. Salesforce is explicit that inference is handled by the provider but the request still routes through the Models API, with Trust Layer features supported.

Three practical constraints from Salesforce's documents:

1. BYOLLM is configured in AI Models (formerly Einstein Studio), which requires Data 360 to be enabled. See Does Agentforce require Data Cloud?
2. The Agentforce Privacy FAQ does not apply to BYO LLM. Your chosen provider is a non-Salesforce application under your own contract.
3. Two charges apply: Salesforce consumption for the action or prompt, plus your provider's own inference charges under your contract.

GPTfy's guide to BYOM in Salesforce covers the AppExchange route, where the model call is made from a Named Credential inside your org.

It depends on the model. Salesforce's Supported Models page draws one line: Anthropic and Amazon models run on Amazon Bedrock entirely within the Salesforce Trust Boundary. Other models run with Salesforce partners, either in a shared trust zone or, with BYOLLM, directly with the provider.

• Anthropic and Amazon on Bedrock. Salesforce states that LLM traffic remains within its private AWS cloud and that customer data is never stored or used for training by external providers. Anthropic is not listed as a sub-processor; AWS is.
• OpenAI. Salesforce's August 2026 sub-processor list names OpenAI (United States) and Microsoft Azure. In a failover, data is temporarily re-routed to an Azure endpoint. Batch processing via Azure OpenAI uses global endpoints and does not support geo-aware routing.
• Google. Listed regions are United States, Germany, Japan, Australia, Brazil, India, and the Netherlands, with US failover.
• Geo-aware routing. Requests go to the nearest data center based on where Data 360 is provisioned for the org.
• Retention. Salesforce says it has zero-data-retention agreements with its third-party LLM partners, and the September 2026 AWS release says Bedrock enforces the same so no provider stores or trains on the data.

Two facts from the same documents are easy to miss. First, data masking in the Einstein Trust Layer is currently disabled for Agentforce agents; Salesforce's Privacy FAQ and its June 2025 blog on LLM data masking cite performance and accuracy as the reasons. Second, when masking is turned on for non-agent use, all models run with a 65,536-token context cap.

Claude. Salesforce's 26 August 2026 Claudeforce announcement states that Claude is the default model for Slack AI, Slackbot, Salesforce in Claude, Headless 360, Agentforce Coworker, and Agentforce Vibes. The specific Claude version is not disclosed.

Salesforce in Claude is different from Claude in Agentforce. Claude in Agentforce runs on Bedrock inside the Salesforce Trust Boundary. Salesforce in Claude is a plugin that runs in the Claude app on any paid Claude plan, signs in as each user's own Salesforce account, and routes actions through Salesforce so business rules apply. It entered beta in September 2026; access requires the current Sales Cloud enterprise edition and Salesforce approval through AgentExchange. Salesforce has not published standalone pricing, and has not stated whether actions triggered from Claude consume Flex Credits.

For agent actions, the rate does not change with the model. Salesforce's Flex Credits rate card dated 17 June 2026 charges 20 credits per standard or custom action and 30 per voice action, with no model dimension on those rows. Sandbox actions are 16 and 24.

For prompt calls outside an agent, the model tier matters. The same rate card lists Basic prompts at 2, Standard at 4, and Advanced at 16 credits per usage unit for Salesforce-enabled models, and BYOLLM starter prompts at 2. Salesforce's developer documentation adds that bringing your own LLM consumes 30% fewer Einstein Requests on the older metering. Which models sit in which tier is on the Help article for billable usage types and changes per release.

How the two consumption models fit together is on Agentforce Flex Credits vs Conversations.

Five model-related announcements from 15 and 16 September 2026, in Salesforce's words:

No pricing change for any model option was announced. Full event coverage is on the Dreamforce 2026 page.

GPTfy is a Salesforce-native managed package. Each AI model is a record in your org that points at a Named Credential you own, so the call to OpenAI, Anthropic, Google, Azure, Llama, DeepSeek, or a private endpoint is made with your key, from your org, under your contract. There is no managed mix and no Data 360 SKU.

Before any prompt leaves the org, GPTfy applies 4-layer masking: field-level, pattern-based, blocklist, and Apex. The masked prompt, the response, and the user are written to a Security Audit Record inside Salesforce. Model selection is per prompt, so a service summary can run on one model and an Apex review on another, with the routing rules held in the org.

If the question is which architecture fits, use the GPTfy vs Agentforce comparison. If the question is which model Agentforce uses, this page is the answer.

Book a demo.