On October 20, 2026, Salesforce plans to reroute requests made to Gemini 2.5 Pro and Gemini 2.5 Flash to Gemini 3.5 Flash. Requests to Gemini 2.5 Flash-Lite will move to Gemini 3.5 Flash-Lite. Salesforce's warning is short but consequential: expected responses can change.

This is not necessarily an outage event. Rerouting is intended to maintain service when an underlying model is retired. The operational risk is subtler: an agent, prompt template, or application can continue returning successful responses while its behavior changes.

For a low-risk drafting assistant, a wording change may be harmless. For a service agent that summarizes a member's case, a healthcare workflow that extracts structured details, an insurance process that classifies an inquiry, or a nonprofit assistant that drafts donor communications, the same change may require review. Organizations using the affected models should treat October 20 as a production change, even if their Salesforce configuration does not otherwise move.

What Salesforce Is Changing

Salesforce says the following requests will be rerouted on October 20:

Current request Rerouted to
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

Google also identifies October 20 as the retirement date for the Gemini 2.5 models in the relevant agent-platform channel. The dates therefore reflect an underlying provider lifecycle event, not merely a Salesforce user-interface change.

The affected scope should be stated carefully. This is not evidence that every Agentforce implementation uses Gemini 2.5. It applies where a Salesforce configuration or custom implementation actually requests one of those models. Salesforce supports model selection across prompt templates, agents, APIs, and some custom actions; other implementations may use a Salesforce-managed default, another named model, or a bring-your-own-model configuration.

Why a Successful Response Can Still Be a Regression

Traditional release checks often ask whether an endpoint responded and whether a transaction completed. Those checks are necessary, but generative AI needs behavioral acceptance criteria as well.

A replacement model may alter:

  • whether an instruction is followed consistently;
  • how a request is classified or routed;
  • the order or choice of agent actions;
  • the accuracy, completeness, or tone of a summary;
  • the shape of JSON or other structured output consumed downstream;
  • citation selection and grounding behavior;
  • latency, prompt tokens, completion tokens, or total consumption;
  • how the system responds to ambiguity, missing data, unsafe requests, or prompt injection.

Not every difference is a defect. A new model may improve an answer. The release decision should be based on business acceptance criteria, not text equality and not the assumption that newer always means equivalent.

A Seven-Step Readiness Plan

1. Prove where Gemini 2.5 is used

Build an inventory before testing. Review AI Models configurations, Prompt Builder templates, agent configuration, custom actions that invoke prompt templates or the Models API, Apex or external integrations that call model APIs, and deployment metadata or technical documentation that names a model API identifier.

Record the asset owner, business process, audience, data sensitivity, downstream actions, current model, environment, volume, and rollback or containment option. Do not infer exposure solely from the fact that Agentforce is enabled.

2. Prioritize by consequence, not prompt count

Score each affected use case by what happens when the response is wrong, incomplete, slow, or differently structured. Put customer-facing, action-taking, regulated-data, high-volume, and machine-consumed outputs first.

For healthcare, include member or patient communications, case summaries, intake extraction, and knowledge-based guidance. For insurers, include claims or policy inquiry classification, correspondence drafting, and escalation. For nonprofit foundations, include applicant or grantee communications, eligibility-support workflows, and donor-facing content. Qualified human review remains essential wherever clinical, coverage, eligibility, legal, or compliance judgment is involved.

3. Freeze a representative baseline

Preserve the current prompt and agent versions, model settings, grounding sources, retrieval configuration, actions, test inputs, and outputs. Include normal cases, edge cases, prior production failures, ambiguous requests, multilingual inputs where relevant, and adversarial or unsafe prompts.

Define measurable acceptance rules before seeing the new output. Useful measures include factual support, required fields, citation validity, action correctness, refusal behavior, escalation, tone, completeness, latency, and token use. Exact-text matching is usually too brittle; business invariants are more useful.

4. Test the replacement model in a sandbox

Salesforce recommends creating a model configuration that matches the rerouted model, using the same hyperparameters where applicable, then retesting prompt templates in Prompt Builder and agents in Agentforce Builder. Testing Center can run batch evaluations for agents and prompt templates.

Use a sandbox. Salesforce notes that agent tests can modify CRM data, and tests consume requests and credits. Use masked, synthetic, or approved test data appropriate to the use case; do not move sensitive production information into a test environment without the organization's established data controls.

5. Compare complete business behavior

Evaluate more than the generated paragraph. If an agent can act, confirm topic or subagent selection, action sequence, parameters, record changes, integration calls, retries, and human handoff. If another system consumes the result, validate schema, null handling, delimiters, length, parsing, and downstream exceptions.

Have operational subject-matter experts review high-consequence samples. Automated scorers help find patterns, but they do not decide whether a coverage explanation, care-related message, or grant communication is appropriate for the organization.

6. Remediate the smallest responsible layer

When a test fails, identify the source before rewriting everything. The issue may be the prompt instruction, grounding data, retriever, model parameters, topic logic, action description, output parser, or an overly broad scope of autonomy.

Prefer a controlled fix that is testable and reversible. Version the prompt or agent, rerun the failed case and the broader regression set, and document the acceptance decision. If the replacement model does not meet the requirement, evaluate another supported model only after confirming feature compatibility, region and trust requirements, limits, cost implications, and organizational approval.

7. Watch the production change as an event

Establish a change window and named owner for October 20. Monitor error rates, response quality samples, escalation and handoff rates, action failures, latency, user feedback, and token consumption. Salesforce's generative AI audit and feedback data can include the model, provider, prompt-template version, parameters, and token counts, subject to feature setup and access controls.

Set thresholds and a containment action for unacceptable behavior. Depending on the implementation, containment could mean disabling an affected feature, routing to a human, switching an approved configuration, or temporarily limiting actions. Retain the evidence needed to explain what changed, what was tested, who approved it, and what happened after rerouting.

The Leadership Decision Before October 20

Executives do not need to review every prompt. They do need a credible answer to four questions:

  1. Do we know which production assets request Gemini 2.5?
  2. Have the highest-consequence workflows passed representative regression tests on the replacement model?
  3. Can we detect a quality, safety, latency, or consumption regression after rerouting?
  4. Is there an owner with authority to contain the problem?

If any answer is no, the organization has a release-readiness gap. Rerouting may keep the service available, but continuity of the endpoint is not continuity of the business outcome.

How Yuniq Can Help

Yuniq's Salesforce services include implementation, customization, integration, testing, and ongoing managed support. For organizations exposed to the Gemini 2.5 reroute, the useful engagement is focused: inventory affected Salesforce AI assets, define business-specific acceptance criteria, execute regression tests in a controlled environment, remediate failures, and establish post-change monitoring.

4. Publishing Notes

Featured-image concept

A clean enterprise release-control scene showing a Salesforce AI workflow moving from “Gemini 2.5” through a dated October 20 change gate to “Gemini 3.5,” with test cases, approvals, and monitoring signals around the gate. Use a restrained healthcare/insurance operations setting rather than a humanoid robot or generic glowing AI image. Do not use Salesforce or Google logos without brand approval.

Suggested social post

Salesforce will reroute Gemini 2.5 model requests on October 20, 2026, and warns that expected responses can change. A successful API response is not proof that the business behavior is unchanged. Our seven-step readiness plan covers exposure discovery, sandbox regression testing, human review, containment, and post-change monitoring for Agentforce and Prompt Builder teams.

Source list

  1. Agentforce and Generative AI Release Notes, Salesforce, Winter '27; reroute item added the week of August 24, 2026. https://help.salesforce.com/s/articleView?id=release-notes.rn_einstein_copilot.htm&language=en_US&type=5
  2. Gemini 2.5 Pro, Google Cloud, last updated October 2, 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/2-5-pro
  3. Gemini 2.5 Flash, Google Cloud, accessed October 6, 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/2-5-flash
  4. Vertex AI Release Notes, Google Cloud, April 2, 2026 announcement updating Gemini 2.5 retirement dates. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/release-notes
  5. Prepare for Model Deprecation and Rerouting, Salesforce Help, accessed October 6, 2026. https://help.salesforce.com/s/articleView?id=sf.generative_ai_model_deprecation_rerouting.htm&language=en_US&type=5
  6. Trust for the Agentic Enterprise, Salesforce Architects, accessed October 6, 2026. https://architect.salesforce.com/docs/architect/well-architected/guide/agentic-enterprise-trust.html
  7. Agentforce Testing Center, Salesforce Help, accessed October 6, 2026. https://help.salesforce.com/s/articleView?id=ai.agent_testing_center.htm&language=en_US&type=5
  8. Einstein Generative AI Audit and Feedback, Salesforce Help, accessed October 6, 2026. https://help.salesforce.com/s/articleView?id=ai.generative_ai_feedback.htm&language=en_US&type=5
  9. Data Model for Generative AI Audit and Feedback, Salesforce Help, accessed October 6, 2026. https://help.salesforce.com/s/articleView?id=sf.generative_ai_feedback_data_model.htm&language=en_US&type=5
  10. Salesforce Services, Yuniq, accessed October 6, 2026. https://yuniq.co/salesforce

Human or subject-matter review before publication

  • Confirm in a Salesforce org that Gemini 3.5 Flash and Gemini 3.5 Flash-Lite are selectable for the relevant licensed features and regions before describing any exact UI path.
  • Have a Salesforce AI architect validate the inventory locations and containment options against Yuniq's typical customer configurations.
  • Have security, privacy, and industry owners review any test-data or audit-data procedure involving PII, PHI, payment, donor, applicant, claims, or policyholder information.
  • Have healthcare, insurance, or nonprofit process owners approve domain-specific examples. The article is operational guidance, not medical, coverage, eligibility, legal, security, or compliance assurance.
  • Recheck the Salesforce release note and Google model lifecycle pages immediately before publication in case the date, target model, availability, or terminology changes.
  • Confirm the preferred CTA wording and landing-page conversion path with Yuniq marketing.