Agentforce LLM choice: default, Claude or your own model
For most manufacturers, the right Agentforce LLM choice is a pinned Salesforce-managed model per subagent, chosen from the few models Salesforce has tested with agents, with a more expensive model only where tests show a quality gap. Salesforce Default is a managed mix you do not control. Bring your own LLM, for example through an existing Azure OpenAI contract, fits prompt templates and custom actions, costs fewer Flex Credits per prompt and moves the data handling into your own contract. Expect new frontier models to reach Agentforce weeks or months after their release.
Which models can an Agentforce agent use today?
The Supported Models page1, which carries no update date, lists three groups:
- Salesforce Default, “a managed mix of trusted models” that currently includes GPT-4o. You change the org default on the Agentforce Agents page in Setup.
- Salesforce-managed models: Claude Haiku 4.5 up to Sonnet 5 and Opus 4.8, Amazon Nova and NVIDIA Nemotron 3 (Beta) on Bedrock; GPT-4o up to GPT 5.6 Luna, Sol and Terra from OpenAI and Azure OpenAI; Gemini 2.5 Flash up to 3.6 Flash and Gemini 3.1 Pro (Beta) on Vertex AI.
- Bring your own LLM (BYOLLM): your own account at Amazon Bedrock, Azure OpenAI, OpenAI or Vertex AI, or any model behind the LLM Open Connector.
Anthropic and Amazon models run on Bedrock “entirely within the Salesforce Trust Boundary”, other managed models “inside a shared trust zone” with Salesforce’s partners. What that means for masking and logging is covered in what the Einstein Trust Layer protects.
The Claudeforce press release of 27 August 20262 names Claude as “a reasoning model for the Atlas Reasoning Engine” and “available in Agent Builder”. It does not call Claude the default for agents, and the developer documentation still names GPT-4o in the default mix. Salesforce’s own CRM reasoning model Koa is, according to the announcement of 15 September 20263, in pilot, with general availability “expected Winter 2026 in U.S. regions”. For a German machine builder planning now, Koa is not yet an option.
Pin the model per agent and subagent
According to the model configuration guide4, every agent uses the org default unless Agent Script sets model_config: for the whole agent, inside a single subagent, or on the agent router that picks the subagent. A subagent model overrides the agent model.
Salesforce names three models as “thoroughly tested with agents”: GPT 4.1, Claude Haiku 4.5 and Gemini 3.5 Flash. For routing, the Agentforce Service agent template uses EinsteinHyperClassifier, a Salesforce model that only classifies and cannot call other tools.
I would pin a model even when the default works. Salesforce can change the mix, and an agent that passed its tests in May may answer differently in November without a change on your side. A pinned model is part of the agent version, so a model change goes through the same gates as any other change, described in testing Agentforce agents before go-live.
What does the model tier cost?
Salesforce assigns every managed model to a usage tier on its model support page5. The rate card updated as of 31 August 20266 prices a prompt by tier and calls prompts to your own model “Starter Prompts”. According to the help article on billable usage7, usage is counted in chunks of 2,000 tokens, rounded up. At the list price of $500 per 100,000 credits8, one Advanced prompt costs $0.08.
2
Flex Credits per prompt on Basic
4
Flex Credits per prompt on Standard
16
Flex Credits per prompt on Advanced
2
Flex Credits per Starter Prompt to your own model
The tested models fall into different tiers: Gemini 3.5 Flash is Basic, Claude Haiku 4.5 and GPT 4.1 are Standard. Every Claude Opus model and GPT 5.6 Sol are Advanced, so moving a subagent from Haiku to Opus multiplies its prompt cost by four.
That help article does not say whether the reasoning engine’s own LLM calls are metered as prompts on top of the 20 credits per action. For prompt templates the tier applies for certain; for reasoning, I would measure it in the pilot with Digital Wallet. The full meter set is in Flex Credits: estimating cost per action, the template side in Prompt Builder or Agentforce agent.
When does bring your own LLM make sense?
BYOLLM connects a model from your own Bedrock, Azure OpenAI, OpenAI or Vertex AI account with your own credentials. The LLM Open Connector9 accepts any model behind a REST endpoint that follows Salesforce’s specification, including one you host. Your team then runs that endpoint. Three points decide whether it is worth it.
Where it is documented. The Supported Models page says custom actions that run prompt templates, Apex or the Models API “can use any Salesforce-managed or BYO model”. The Agent Script guide only refers to models on the supported list and says nothing about BYO models for reasoning.
Who is responsible for the data. Trust Layer features are “fully supported” for BYOLLM, according to the Supported Models page. The contract side changes. Salesforce’s Agentforce privacy FAQ10 (June 2025) states that its answers “do not apply to BYO LLM”, that the provider is “a Non-SFDC Application” and that the customer must comply “with its own contract with the LLM provider”. Retention, training use and region then follow your Azure or AWS agreement. If your IT has already negotiated that agreement, this can be the reason to use BYOLLM; for everyone else it is extra contract work. That is my reading of the documents, not legal advice.
What it saves. Starter prompts cost 2 credits instead of 4 on Standard, and the Supported Models page speaks of “30% fewer Einstein Requests”. Your provider bills the tokens on top.
The model list lags behind the vendors
Anthropic’s model overview11 lists Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 as current models; Salesforce’s managed list ends at Sonnet 5 and Opus 4.8. Even Salesforce’s own pages differ: the help page lists Gemini 3.7 and 3.8 Flash, the developer page stops at 3.6 Flash. I assume a new frontier model reaches Agentforce weeks to months after release and never build a business case on a model that is not on the list.
Models also leave. After its shutdown date a deprecated model is rerouted to a replacement, for example Claude Sonnet 4 to Sonnet 4.6, and Gemini 3 Pro (Beta) was retired on 23 April 2026. The model page lists deprecation and reroute dates, and I found no stated notice period. A reroute is a model change your tests have not seen.
Region is the third check. Managed models “eventually re-route to a US region” when a request fails in the designated region, unless an admin turns on “Enable In-Region Model Requests Only”, which covers “some models”. The help page lists eu inference profiles for Claude Sonnet 5 and Opus 4.8, and only the United States for GPT 5.6 Terra. More in Agentforce EU data residency.
Worked example: a service agent with an Azure OpenAI contract
Assume a machine builder with Service Cloud, Field Service, a dealer portal and SAP. Corporate IT has an Azure OpenAI contract and asks why Agentforce should not use it. All numbers are assumptions.
- Router: EinsteinHyperClassifier, as in the service template.
- Order status for dealers from SAP via MuleSoft: Gemini 3.5 Flash (Basic), because the answer is a lookup.
- Warranty pre-check, the rule itself in Agent Script logic: Claude Haiku 4.5 (Standard) reads serial number and fault description from free text.
- Troubleshooting from manuals for dealer technicians: Claude Haiku 4.5, tested against Opus 4.8 on 100 real fault descriptions before anyone pays for Advanced.
A prompt template also writes a case summary for the field service technician: 3,000 a month at about 5,000 tokens, three prompts each.
| Option for the summary | Credits per month | Salesforce side |
|---|---|---|
| Managed Standard model | 3,000 × 3 × 4 = 36,000 | $180 |
| BYOLLM via Azure OpenAI (Starter) | 3,000 × 3 × 2 = 18,000 | $90 plus Azure tokens |
| Managed Advanced model | 3,000 × 3 × 16 = 144,000 | $720 |
BYOLLM saves at most $90 a month, about $1,080 a year, before the Azure token bill. That alone does not justify a second contract path; the model tier moves four times as much. Here I would recommend managed models for the agent, BYOLLM only if the Azure agreement is IT’s approved data path for service data, and Opus for troubleshooting only if the test shows a clear gap.
Pin and document the model per subagent
- Pin the model per agent and subagent in Agent Script.
- Start with a tested model in the lowest tier that passes; upgrade a subagent only with test evidence.
- Check the region of each model and decide on in-region only.
- Use BYOLLM with a reason: an approved contract, a model Salesforce does not offer, or data rules that require your own provider.
- Check the model page and release notes regularly and rerun the tests after every reroute.
- Measure in Digital Wallet whether reasoning calls show up as prompts.
I treat these decisions as Salesforce architecture work, next to data model and permissions. For each subagent I write down model, tier, region and test result, because every change touches cost, contracts and tests.
Sources11
- 1Supported Models page (opens in a new tab)developer.salesforce.com
- 2Claudeforce press release of 27 August 2026 (opens in a new tab)salesforce.com
- 3announcement of 15 September 2026 (opens in a new tab)investor.salesforce.com
- 4model configuration guide (opens in a new tab)developer.salesforce.com
- 5model support page (opens in a new tab)help.salesforce.com
- 6rate card updated as of 31 August 2026 (opens in a new tab)salesforce.com
- 7help article on billable usage (opens in a new tab)help.salesforce.com
- 8list price of $500 per 100,000 credits (opens in a new tab)salesforce.com
- 9LLM Open Connector (opens in a new tab)developer.salesforce.com
- 10Agentforce privacy FAQ (opens in a new tab)salesforce.com
- 11model overview (opens in a new tab)platform.claude.com
Frequently asked questions
Which LLM does Agentforce use by default?
The default option is Salesforce Default, which the Supported Models page describes as a managed mix of trusted models that currently includes GPT-4o. Salesforce decides what is in the mix, so pin a model if you need a fixed one.
Can I use a different model for each subagent?
Yes. In Agent Script you set model_config at agent level, inside a subagent or on the agent router. A subagent model overrides the agent model, which overrides the org default set in Setup.
Can I use my existing Azure OpenAI contract with Agentforce?
Yes, through bring your own LLM, which supports Amazon Bedrock, Azure OpenAI, OpenAI and Vertex AI with your own credentials. Salesforce documents it for prompt templates, the Models API and custom actions; confirm in writing before you plan it as the agent's reasoning model.
Is a bring your own LLM model cheaper in Flex Credits?
Per prompt, yes: Starter prompts for your own model cost 2 credits per 2,000 tokens, against 4 on a Standard managed model. Your provider then bills the tokens under your own contract, so the total is rarely much lower.
When will Claude Opus 5.5 be available in Agentforce?
Salesforce has not published a date. As of 27 October 2026 the Supported Models page lists Claude up to Sonnet 5 and Opus 4.8, although Anthropic already lists Opus 5.5 and Sonnet 5.5 as current models.