Gemini 3.5 Flash vs Gemini 3 Flash Preview: Migration Guide
This is not only a model-ID swap. Google documents SDK, thinking, function-response, and prompt-behavior changes that can alter production results. On EvoLink, keep the gateway integration stable while replaying the same workload through both model routes.
Gemini 3.5 Flash vs 3 Flash: Quick Decision
| Your situation | Better first route | Why |
|---|---|---|
| New production application | Gemini 3.5 Flash | GA, stable, and recommended by Google for migration |
| Existing Preview workflow that already passes | Replay both | Avoid paying more without a measured benefit |
| Agentic coding or multi-step tool use | Gemini 3.5 Flash | Google's current Flash model for agentic and coding tasks |
| High-volume extraction or translation | Gemini 3.1 Flash-Lite | Stable cost-first route without long-term Preview dependence |
| Preview testing or behavior research | Gemini 3 Flash Preview | It remains available for continued testing |
| Migration with no prompt or tool evals | Stay temporarily | Behavior changes can break acceptance even when requests still run |
Google recommends migrating to 3.5 Flash for GA stability and improved reasoning, but it has not listed a shutdown date for Gemini 3 Flash Preview in the current migration guide.
Verified Comparison
| Dimension | Gemini 3.5 Flash | Gemini 3 Flash Preview | Production meaning |
|---|---|---|---|
| Model ID | gemini-3.5-flash | gemini-3-flash-preview | Change the model route during migration |
| Status | Stable, GA | Preview | 3.5 is the safer long-term default |
| Google standard input price | $1.50 / MTok | $0.50 / MTok | 3.5 is 3× the standard input price |
| Google standard output price | $9.00 / MTok | $3.00 / MTok | Thinking tokens are included in output billing |
| Context caching | $0.15 / MTok | $0.05 / MTok | Same 3× standard-price relationship |
| Context window | 1M tokens | 1M-token class | No migration decision should rely on a context jump |
| Maximum output | 65K tokens | 65K tokens | Same documented output class |
| Default thinking level | medium | high | The same prompt can change latency, cost, and depth |
| Multimodal input | Text, image, video, audio, PDF | Text, image, video, audio, PDF | Re-test media-heavy prompts rather than assuming parity |
| Function calling | Supported | Supported | Response matching rules require attention |
| Structured output | Supported | Supported | Validate schemas on replayed traffic |
| Computer Use | Preview support | Preview support | Keep preview-tool risk separate from model GA status |
The Migration Is More Than a Model-ID Swap
The minimum route change is simple:
gemini-3-flash-preview → gemini-3.5-flashBut Google's migration guidance also calls for the following checks:
| Migration item | Current guidance | Why it matters |
|---|---|---|
| Google Gen AI SDK | Use google-genai v2.0.0 or later for the current Interactions API | The May 2026 Interactions API update includes breaking changes |
| Sampling controls | Remove custom temperature, top_p, and top_k | Gemini 3.x is optimized for defaults |
| Thinking control | Prefer thinking_level over thinking_budget | Using both returns a 400 error |
| Default effort | 3.5 defaults to medium, Preview to high | Quality, latency, and tool-call behavior can change |
| Function responses | Match id, name, and response count | Mismatches can error or produce empty responses |
| Thought preservation | Reasoning context carries across turns | Multi-turn quality may improve, but token use can change |
| Prompt replay | Re-test existing prompts | A compatible request can still produce different behavior |
| Tool budget | Lower thinking or constrain tools if calls increase | Tool overuse can raise latency and cost |
If you call Gemini through EvoLink's OpenAI-compatible interface, the Google SDK upgrade is not necessarily part of your application path. If you use native Gemini or Interactions API requests through EvoLink, follow Google's SDK and request-shape migration guidance. In both cases, the behavioral replay still matters.
Thinking Behavior Is the Most Important Default Change
high thinking. Gemini 3.5 Flash defaults to medium.That means a direct model-ID swap may produce:
- lower latency on some tasks
- fewer reasoning tokens
- fewer or different tool calls
- different answers on difficult code or math
- a lower task cost than a naive 3× token-price estimate
- a quality regression if a Preview prompt depended on
high
medium for the migration baseline. Use low when speed and tool-call restraint matter, and test high only for the hardest reasoning or coding work. Do not compare Preview at high with 3.5 at medium and call the result a pure model comparison; record the thinking level with every evaluation run.Function Calling Can Fail Even When the Model Is Available
Google's current rules require every function response to match the preceding function call:
- include the call
id - use the same function
name - return exactly one response for each call
- put multimodal function content inside the function response
finish_reason: STOP. A migration test that checks only HTTP success can therefore miss a broken agent loop.For production evaluation, count:
| Agent signal | What failure looks like |
|---|---|
| Function-call completion | Missing or unmatched function result |
| Empty-response rate | Successful request with no usable output |
| Tool-call count | More calls than the task requires |
| Loop completion | Agent never reaches the accepted end state |
| Schema validity | Output fails downstream validation |
When to Migrate to Gemini 3.5 Flash
You need a stable production model
Gemini 3.5 Flash is GA and stable. That is the strongest reason to use it for a new product path instead of building permanent production dependencies on a Preview route.
Your workload is agentic or long-horizon
Google positions 3.5 Flash for sustained agentic execution, coding, sub-agent deployment, and multi-step work. These tasks also expose the behavioral changes most clearly, so use a controlled replay rather than a one-prompt demo.
Preview retries are raising effective cost
The standard token price is higher, but cost per successful task can improve if 3.5 reduces retries, tool loops, or escalation to a larger model.
You want a durable three-tier Gemini policy
On EvoLink, a practical family policy is:
- Gemini 3.1 Flash-Lite for stable high-volume, cost-first work
- Gemini 3.5 Flash for balanced production reasoning and agents
- a Pro route for the hardest capability-sensitive tasks
That is more durable than keeping Preview as the permanent middle or budget tier.
When to Keep Gemini 3 Flash Preview Temporarily
Do not migrate solely because 3.5 is newer if:
- Preview already meets the exact acceptance rubric
- a 3× standard token-price increase is not justified by task outcomes
- the workload is temporary or explicitly experimental
- the team cannot yet test function-response and multi-turn behavior
- changing the model during a release freeze creates more risk than value
Keeping Preview should have an owner and review date. It should not become an unmonitored default just because no shutdown date has been announced.
Cost: Compare the Same Task, Not Only the Price Table
At Google standard pricing, a workload using 10M input and 2M output tokens would look like this before caching, search grounding, retries, or discounts:
| Model | Input cost | Output cost | Total |
|---|---|---|---|
| Gemini 3 Flash Preview | $5 | $6 | $11 |
| Gemini 3.5 Flash | $15 | $18 | $33 |
That is a useful budget ceiling, not the final migration answer. Track:
cost per successful task =
(tokens + retries + tool calls + fallbacks + review cost) / accepted tasksA Production Replay Plan
Build a matched set from real traffic rather than benchmark prompts selected after seeing the output.
| Workload | Keep constant | Measure |
|---|---|---|
| Structured extraction | Schema, input, validation rules | Valid-output and retry rate |
| Multi-step coding | Repository state, tools, stop condition | Accepted change, tests, latency |
| Function-call chain | Functions and injected results | Match rate, empty responses, completion |
| Long document analysis | Document and question set | Coverage, citations, token use |
| Cost-sensitive batch | Input mix and acceptance rules | Cost per successful item |
medium. Add a separate 3.5 high arm only when the workload needs it. That makes thinking configuration visible instead of hiding it inside the model name.
Recommended EvoLink Migration
| Stage | Route policy | Pass condition |
|---|---|---|
| 1. Inventory | Record interface, model ID, thinking config, tools, and prompts | No unknown production path remains |
| 2. Shadow replay | Run saved Preview tasks on 3.5 | Request and function flows complete |
| 3. Limited traffic | Route a small production cohort to 3.5 | Quality, latency, and cost stay in bounds |
| 4. Workload promotion | Make 3.5 default for the task classes it wins | Preview rollback remains available |
| 5. Route cleanup | Move cost-first work to Flash-Lite and remove unnecessary Preview use | Ownership and monitoring are documented |
EvoLink's value in this migration is operational: keep one API gateway and one billing surface while comparing model routes. Change the model selection without turning the evaluation into a provider-integration project.
Common Migration Mistakes
- Treating a model-ID swap as proof that agent loops still work
- Comparing Preview
highwith 3.5mediumwithout recording the difference - Using both
thinking_budgetandthinking_level - Ignoring empty responses from mismatched function results
- Copying Google standard pricing into an EvoLink bill estimate without checking the live route price
- Keeping Preview forever because no retirement date is listed
- Sending every cheap task to 3.5 instead of evaluating stable Flash-Lite
Continue on EvoLink
- Check live route details on Gemini 3.5 Flash.
- Keep Gemini 3 Flash Preview as a temporary baseline during replay.
- Compare family routes on the Gemini API Family page.
- Use the Gemini 3.5 Flash pricing guide for detailed budgeting.
- Use Gemini 3.5 Flash for coding agents for the coding-specific evaluation path.
Sources
- Google: What's new in Gemini 3.5 Flash
- Google: Gemini Developer API pricing
- Google: Gemini models
- Google: Gemini API changelog
- Google: Interactions API breaking changes
FAQ
Should I migrate from Gemini 3 Flash Preview to Gemini 3.5 Flash?
Yes for new production workloads and for existing workloads where GA stability, reasoning, or agent execution justifies the higher standard price. Replay existing traffic before making 3.5 the default everywhere.
Is Gemini 3 Flash Preview still available?
Yes. Google currently says it remains available for developers who want to continue testing. The current migration guide does not list a shutdown date.
Is the migration only a model-ID change?
gemini-3-flash-preview to gemini-3.5-flash, but Google also recommends SDK, thinking-control, function-response, and prompt-behavior checks.Why can the same prompt behave differently after migration?
medium thinking while Gemini 3 Flash Preview defaults to high. Thought preservation and tool behavior can also change multi-turn results.Can I continue using thinking_budget?
thinking_level. Do not send both in the same request because that returns a 400 error.Does Gemini 3.5 Flash cost three times more?
At Google's standard paid-tier pricing, its input, output, and cache rates are 3× Preview. EvoLink route pricing may differ, and successful-task cost also depends on retries, thinking, tools, caching, and fallbacks.
Is Gemini 3.1 Flash-Lite still a preview model?
No. Google lists Gemini 3.1 Flash-Lite as stable and positions it for high-volume, cost-efficient tasks. It is the better long-term cost-first comparison than relying on Gemini 3 Flash Preview indefinitely.
How should EvoLink users test the migration?
Replay the same prompts, tools, thinking settings, schemas, and stop conditions through both routes. Compare accepted-task rate, empty responses, tool calls, latency, token use, and total task cost before promoting 3.5.
