Skip to main content

Google: Migrating to Gemini 3 and 3.1

Use this guide when you are moving older Gemini traffic onto:
  • google/gemini-3-pro-preview
  • google/gemini-3-flash-preview
  • google/gemini-3.1-pro-preview
  • google/gemini-3.1-flash-lite-preview
Gemini 3 introduced real API-level changes around reasoning control and multimodal handling. Gemini 3.1 then raised the baseline again, especially for more complex reasoning-heavy tasks.

What changed

  • Gemini 3 introduced thinking_level as the main control for reasoning depth
  • Gemini 3 added media_resolution for multimodal vision handling
  • Gemini 3 uses Thought Signatures to preserve reasoning context across calls
  • in stricter workflows, missing Thought Signatures can produce degraded quality or 400 errors
  • Gemini 3.1 Pro is positioned as the stronger reasoning upgrade for complex tasks and agentic workflows

What to change in your integration

1. Re-map reasoning controls

If your app exposes a generic reasoning knob, map it carefully for Gemini 3 routes. In OpenAI-compatible flows, reasoning_effort maps conceptually to Gemini thinking_level, but the behavior is not one-to-one and needs retesting.

2. Handle Thought Signatures correctly

If you use the official Gemini SDKs and standard chat history, signatures are usually handled for you. If you build lower-level request flows, especially around function calling or image generation and editing, make sure the returned signatures are preserved and sent back exactly as required.

3. Revisit prompt style

Gemini 3 guidance favors more direct prompts and parameter-driven control rather than longer prompt scaffolding. If your current prompts are heavily compensating for older model behavior, simplify them and re-evaluate.

4. Tune multimodal settings

If your routes include images, documents, or mixed media, re-test media_resolution settings instead of assuming older defaults still fit.

What to test

Prompt and reasoning behavior

  • concise prompts versus heavily scaffolded prompts
  • step-by-step reasoning tasks
  • long-context synthesis
  • agent-style workflows that depend on maintaining reasoning continuity

Response handling

  • signature preservation across turns
  • function-calling success rate
  • schema pass rate
  • empty, partial, or malformed outputs after changing thinking controls

Production metrics

  • latency at each thinking_level
  • token usage and cost
  • multimodal route performance under different media_resolution settings
  • task success when comparing Gemini 3 versus Gemini 3.1 Pro

Rollout advice

Treat Gemini 3 and Gemini 3.1 as separate migrations.
  1. First migrate the request shape and signature handling.
  2. Then re-run prompt and schema evaluations on Gemini 3.
  3. Promote Gemini 3.1 Pro separately for the routes that need deeper reasoning.
  4. Keep a simpler or cheaper fallback path for high-volume workloads.

Sources

Last modified on March 11, 2026