Salesforce Is Rerouting Gemini 2.5 on October 20: Find What Calls It and Test Before the Switch

On October 20, 2026, Salesforce will reroute requests for Gemini 2.5 Pro and Gemini 2.5 Flash to Gemini 3.5 Flash, and requests for Gemini 2.5 Flash-Lite to Gemini 3.5 Flash-Lite. It isn't an outage. Calls keep working, but Salesforce warns that expected responses can change, and recommends testing your prompts and applications with the new model in Prompt Builder and AI Models as soon as possible.

It only matters where something in your org actually asks for a Gemini 2.5 model. If that's you, you have a few working days. Here's how to use them.

What changes on October 20

| If something requests… | From Oct 20 it gets… | |---|---| | Gemini 2.5 Pro | Gemini 3.5 Flash | | Gemini 2.5 Flash | Gemini 3.5 Flash | | Gemini 2.5 Flash-Lite | Gemini 3.5 Flash-Lite |

Note the first row. An earlier preview of this release note said Gemini 2.5 Pro would move to Gemini 3.5 Pro. The current Salesforce note says Gemini 3.5 Flash. If you planned around "Pro goes to Pro," plan again.

Salesforce reroutes a model when the provider retires it. Google lists the same date, October 20, 2026, for retiring Gemini 2.5 Pro, Flash and Flash-Lite on its enterprise agent platform. So this date comes from the provider and Salesforce can't simply push it back.

What doesn't change: your templates, agents and integrations keep returning successful responses. That's the risk. Something can keep "working" while its behavior shifts: a summary gets longer, a classification lands in a different bucket, a JSON field comes back empty, an agent picks a different action.

Does this affect my org?

Only if something requests a Gemini 2.5 model. Having Agentforce turned on doesn't mean you're affected. Many setups use a Salesforce-managed default, a different named model, or a model you brought yourself.

Salesforce's release note scopes the change to Lightning Experience in Enterprise, Performance and Unlimited editions with an Einstein for Sales, Einstein for Platform or Einstein for Service add-on.

If you can't answer "what in our org asks for Gemini 2.5?" today, start with Step 1.

Step 1: Find everything that asks for Gemini 2.5

| Where to look | What to record | |---|---| | AI Models configurations (in Einstein Studio) | Which configured models are built on a Gemini 2.5 foundation model, and their settings | | Prompt Builder templates | The model each template uses | | Agents in Agentforce Builder | Any model setting on the agent, and which templates and actions it calls | | Custom actions | Actions that invoke a prompt template or call a model directly | | Apex and external integrations | Code that calls a model API with a named model | | Your repo, deployment metadata and runbooks | Any place a Gemini 2.5 model name is hard-coded |

For each asset, write down the owner, the business process, who sees the output (internal staff or customers), how sensitive the data is, what happens downstream, the environment, and how you'd switch it off if it misbehaves.

Step 2: Rank by consequence, not by count

Twenty internal drafting helpers can matter less than one template that writes back to records. Put these first:

**Customer-facing output:** case replies, email drafts that go to customers, portal answers.
**Output that takes an action:** updates records, sends messages, calls an integration.
**Output a machine reads:** JSON or field values that a flow, Apex class or integration parses.
**Sensitive or regulated data** in the prompt or the grounding.
**High-volume** use, where a small change repeats thousands of times.

Typical mid-market examples: a case summary on the service console, lead or opportunity fields generated from activity, a classifier that routes incoming cases, a renewal email draft.

Step 3: Freeze a baseline before you change anything

For each high-consequence asset, save:

The current prompt template and agent versions, model settings and grounding.
A set of test inputs: normal cases, edge cases, ones that broke before, ambiguous ones, and a few that try to push the agent off-script.
The current outputs for those inputs.

Then write your pass rules before you look at new output. "Required fields present," "right action chosen," "escalates when it should," "tone acceptable," "parser doesn't fail." Don't expect an exact text match. A new model will word things differently even when it's doing the job right.

Step 4: Test the replacement model in a sandbox

Salesforce's guidance for rerouting is to create a new model configuration in Einstein Studio that matches the rerouted model, with the same settings as the retired one, and then retest your prompt templates against it in Prompt Builder. Retest agents in Agentforce Builder the same way.

Do this in a sandbox. Agent tests can act on CRM data. Use masked, synthetic or approved test data, not a copy of sensitive production records your data rules don't allow in a sandbox.

Step 5: Compare behavior, not paragraphs

For agents, check:

Did it pick the same topic and the same action?
Did it pass the same parameters?
Did it change the same records, call the same integration, hand off to a human at the same point?

For structured output, check the schema, null fields, length limits, and whether anything downstream throws a parse error.

Step 6: Fix the smallest layer

When a test fails, change the smallest thing that fixes it: one instruction in the prompt, the grounding, a model setting, an action description, or the parser downstream. Version the change, rerun the failed cases and then the full set, and write down what you decided.

You can also choose a different supported model on purpose instead of taking the reroute target. If you do, it gets the same test run and the same sign-off.

Step 7: Treat October 20 as a change event

Name one owner for the day.
Decide what you'll watch: errors, a sample of outputs, escalation and handoff rates, action failures, response time, user complaints.
Set a threshold and a containment step in advance: turn off the feature, route to a human, or limit what the agent can do.
Keep the evidence: baseline, test results, decisions. You'll want it for the next reroute.

If you're also still working out who can build agents and what they can reach, that's a broader inventory, and a separate job from this regression test.

Four questions to answer before Monday, October 19

Do we know everything in production that requests Gemini 2.5?
Have the high-consequence ones passed against the replacement model?
Will we notice a regression after October 20?
Who can switch it off, and how fast?

Pre-October 20 checklist (copy this table)

| Org | Asset (template / agent / action / Apex) | Current model | Replacement | Owner | Consequence tier | Baseline saved? | Sandbox result | Fix applied | Oct 20 monitor + containment | |---|---|---|---|---|---|---|---|---|---| | | | | | | | | | | |

FAQ

What is Salesforce changing on October 20, 2026?

Requests for Gemini 2.5 Pro and Gemini 2.5 Flash are rerouted to Gemini 3.5 Flash, and requests for Gemini 2.5 Flash-Lite are rerouted to Gemini 3.5 Flash-Lite.

Is Gemini 2.5 Pro being rerouted to Gemini 3.5 Pro?

No. An earlier preview said so, but the current Salesforce release note says Gemini 2.5 Pro goes to Gemini 3.5 Flash.

Will my Agentforce agents stop working?

That's not expected. Requests are rerouted, not dropped. But Salesforce warns that expected responses can change, so behavior can shift while calls keep succeeding.

Does this affect every org with Agentforce?

No. Only templates, agents, actions and integrations that request a Gemini 2.5 model.

Where do I test the new model?

Salesforce recommends Prompt Builder and AI Models: set up a model configuration matching the rerouted model, then retest your prompt templates. Retest agents in Agentforce Builder. Use a sandbox.

What should I monitor after October 20?

Errors, a sample of outputs, escalation and handoff rates, failed actions, response time and user feedback. Set a threshold and a containment step before the day.

While you're in there

A model reroute is a good reminder of how much in a Salesforce org changes underneath you. While you're inventorying AI assets this week, take two minutes to see where the rest of your org's health stands against peers. Three answers give you a score on the page, with no account and no org connection. It's an estimate from your answers, not a test of your AI setup.

*[Start the free health check →](https://wrenk.ai/free-health-check?utm_source=wrenk_insights&utm_medium=blog&utm_campaign=wrenk-12-gemini-reroute)*