Prepare Gemini 2.5 apps for Google’s December 1 model switch

Google announced on November 2, 2026 that applications using Gemini 2.5 Pro or Gemini 2.5 Flash through its Developer API will be switched to replacement models on December 1. The scheduled change affects businesses and developers running assistants, document-processing tools and automations on those models.

Automatic routing does not mean an application will behave unchanged. Some old settings will stop influencing answers; other request formats will fail. Preparing now means checking complete business tasks, not merely whether the service sends back a response.

Find the applications that need to move

Google’s release notice specifies these destinations, with replacement-model prices applying from the switch:

The notice says both retiring models have already left AI Studio’s model selector; that is separate from the December API change. It does not include Gemini 2.5 Flash-Lite, Live/audio or image variants. Nor does it establish a deadline for consumer Gemini subscriptions, Google Cloud endpoints or third-party providers.

Check model identifiers in application settings, scheduled jobs and automation connectors. If a supplier runs the integration, ask for the affected-workflow list and a named migration owner. Configure the chosen replacement explicitly before December 1. Google has not specified the cutover’s time of day or how long the old identifiers will remain accepted afterward.

Repair request settings before moving live traffic

Inspect the request actually sent to Google, including defaults added by integration libraries. The important distinction is between a setting that silently stops working and a request the API rejects:

• temperature, top_p and top_k: Google’s parameter-change guide says these sampling controls are ignored on the affected newer models, with rejection planned for future generations. Remove them; an accepted request no longer proves your old output controls are active.

• Prefilled answers: a request whose last non-empty turn has the model role returns HTTP 400, according to the same guide. Instead of starting the model’s answer for it, put formatting requirements in system instructions or use a supported structured-output schema.

• Thinking and candidate controls: Google’s migration checklist says to replace thinking_budget with thinking_level and remove candidate_count, which Gemini 3 and later do not support. Follow the field naming used by your SDK; a numeric budget is not an exact equivalent of a named effort level.

Choose thinking effort separately for each replacement. Google lists Gemini 3.8 Flash’s default as medium and Gemini 3.5 Flash-Lite’s as minimal. Both offer low, medium and high, but 3.8 Flash rejects minimal. A configuration shared between the two therefore needs a model-specific check.

Check conversations and tool results, not just first answers

Apps that call tools need to preserve the model’s returned thought signatures—opaque data used to carry reasoning context between steps. Google’s GenerateContent signature guidance says missing required function-call signatures can cause HTTP 400 on Gemini 3. Keep the original response parts and their ordering; do not rebuild conversation history from readable text alone.

Tool-result fields also depend on the endpoint. Interactions examples use call_id and name; the GenerateContent reference and Google’s ADK implementation use id, name and response. The 3.8 migration checklist instead says call_id for GenerateContent. Check your endpoint’s reference and SDK schema rather than treating those field names as interchangeable.

If prompts demand structured text immediately before a tool call, Google documents possible Malformed_Function_Call errors. Its workarounds include a separate update function or removing that pre-tool-text requirement.

An endpoint redesign is not required for this deadline: Google says GenerateContent remains supported. If you choose Interactions with store=true and previous_interaction_id, the server manages signatures. Stateless Interactions clients must preserve complete thought blocks and signed built-in-tool results themselves.

That storage choice has consequences. Interactions retains data for 55 days on paid accounts by default, with shorter configurable windows, and one day on free accounts. Setting store=false prevents subsequent use of previous_interaction_id. Review the retention controls before adopting server-managed history.

Budget for January as well as December

These Standard paid rates are US dollars per million tokens, the units of model input and output. Billed output includes thinking.

Illustrative calculation: hold monthly usage at 20 million input tokens and one million billed output tokens, with every prompt at or below 200,000 tokens. Exclude caching, tools, discounts and tax:

• 2.5 Pro: 20 × $1.25 + 1 × $10 = $35.

• 3.8 Flash in December: 20 × $0.75 + 1 × $3.75 = $18.75.

• 3.8 Flash in January: 20 × $1.50 + 1 × $7.50 = $37.50.

January doubles December’s replacement rates, not necessarily the old bill. Actual token use and retries may change; Flash’s matching text rates likewise do not guarantee matching invoices.

Make the rollout depend on completed tasks

Before changing production, run representative cases in a staging environment. For document extraction, check required fields and usable formatting. For assistants, include follow-up questions, refused requests, tool failures and completed multi-step tasks. Use non-sensitive examples and simulated tools where real actions would affect customers or business systems.

Compare acceptable results, response time, billed tokens and retries against the current workflow. Roll out to a limited share of traffic first, preserving existing action permissions and human review. Decide who can pause the rollout and which supported alternative—or manual process—will handle essential work.

Do not rely on switching back to the old model name as a rollback plan: once Google’s announced routing applies, that name points to the replacement. Validate a genuinely separate fallback before the deadline.

Teams managing several providers can also check the separate Claude Opus 4.8 fast-mode retirement guide. Keep each provider’s failure behavior and migration requirements separate.