Gemini 3.6 Flash 與 Gemini 3.5 Flash-Lite 詳解：您應該使用哪一個？

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite at a Glance

Google released both models on July 21, 2026. Unlike many recent Gemini launches, neither model entered the API as a temporary preview. Google lists both as stable and generally available for production use.

Both models have a March 2026 knowledge cutoff, according to their official model cards. For more recent information, developers need to use search grounding or provide current data in the prompt.

What Is Gemini 3.6 Flash?

Gemini 3.6 Flash is Google’s latest general-purpose Flash model for coding, knowledge work, multimodal analysis, and multi-step agentic execution.

It is based on Gemini 3.5 Flash rather than being an entirely separate generation of the Gemini architecture. The main goal of the update is to make Flash more efficient during real work: fewer unnecessary output tokens, fewer reasoning steps, fewer tool calls, and fewer repeated execution loops.

That distinction matters. Gemini 3.6 Flash is not simply trying to produce a higher score on every intelligence benchmark. It is trying to complete the same or better work with less computation and less waiting.

According to Google’s launch announcement, Gemini 3.6 Flash uses approximately 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. Google also reports that it takes fewer turns and tool calls to complete multi-step workflows.

Official evaluations show improvements in several practical areas:

• DeepSWE increased from 37% to 49%, suggesting fewer incorrect code edits and execution loops.
• MLE-Bench increased from 49.7% to 63.9% for machine learning research tasks.
• OSWorld-Verified increased from 78.4% to 83% for computer-use tasks.
• GDPval-AA v2 increased from 1349 to 1421 for knowledge-work performance.

These are vendor-reported results, so they should not be treated as universal guarantees. However, they support Google’s positioning of Gemini 3.6 Flash as a more reliable model for code migrations, technical diagnostics, document analysis, chart interpretation, and tool-using agents.

The model accepts text, images, video, audio, and PDF files, but it only generates text. It does not directly generate images, videos, or speech.

What Is Gemini 3.5 Flash-Lite—and What Does GA Mean?

Gemini 3.5 Flash-Lite is Google’s fastest and least expensive model in the Gemini 3.5 family. It is optimized for low-latency, high-volume tasks where throughput and API cost matter more than maximum reasoning quality.

The “GA” in Flash-Lite GA means Generally Available. It is an availability status, not part of the model’s name. A GA model is intended for stable production use rather than short-term preview testing.

Gemini 3.5 Flash-Lite is based on Gemini 3.1 Flash-Lite—not Gemini 3.6 Flash. Therefore, it should not be understood as a compressed version of the new 3.6 model. The two releases come from different upgrade paths:

• Gemini 3.5 Flash → Gemini 3.6 Flash
• Gemini 3.1 Flash-Lite → Gemini 3.5 Flash-Lite

Google positions Flash-Lite for use cases such as:

• Document classification and structured extraction
• Receipt and invoice processing
• Product attribute extraction
• Translation and localization
• Search-result processing
• High-volume customer-support routing
• Tabular data processing
• Repetitive tool calls
• Parallel subagent execution

Independent testing cited in Google’s announcement measured Gemini 3.5 Flash-Lite at approximately 350 output tokens per second. Actual API performance will vary with prompt length, thinking level, server load, tools, and region, but the result illustrates the model’s throughput-focused design.

Flash-Lite is also considerably more capable than its predecessor. Google reports improvements from 31% to 54% on Terminal-Bench 2.1 and from 60.1% to 72.2% on its long-context evaluation.

It can use higher thinking levels for more complicated tasks, but doing so changes its cost and latency profile. If every request requires deep reasoning, Flash-Lite may lose some of the economic advantage that makes it attractive.

What Do “Flash” and “Flash-Lite” Mean in Gemini?

“Flash” is not an official acronym. It is Google’s product label for Gemini models that prioritize speed, efficiency, and scalable inference.

The practical meaning of Flash is:

• Faster than heavier flagship models
• Less expensive to run at scale
• Suitable for interactive applications
• Still capable of reasoning, coding, and multimodal understanding

Flash-Lite moves further toward the efficiency end of that spectrum. It is designed for workloads with a large number of relatively bounded tasks, such as extracting fields from thousands of documents or translating millions of short content items.

“Lite” does not mean that the model cannot reason or process multimodal inputs. Gemini 3.5 Flash-Lite still supports thinking, function calling, structured output, search grounding, and a 1M-token context window.

The difference is how Google expects each model to be deployed:

• Flash: Choose it when the model must understand a complicated goal and decide what to do.
• Flash-Lite: Choose it when the task is already defined and needs to be completed quickly at scale.

Gemini 3.6 Flash vs Gemini 3.5 Flash: What Actually Improved?

Gemini 3.6 Flash is the direct successor to Gemini 3.5 Flash, but “successor” does not mean that every intelligence score has increased.

The most meaningful improvements are efficiency, coding reliability, tool use, and total time per task.

In Artificial Analysis testing, both models scored 50 on its Intelligence Index. That makes it difficult to argue that Gemini 3.6 Flash represents a large increase in general intelligence.

However, Gemini 3.6 Flash completed the tested tasks in less than half the average time. Its output price is also approximately 16.7% lower, while the model uses fewer output tokens in many agentic workloads.

For production users, this may be more valuable than a small benchmark increase. A model that reaches a similar answer with fewer tokens, fewer failed tool calls, and fewer correction loops can reduce both the infrastructure cost and the time users spend waiting.

There is little pricing incentive to begin a new deployment on Gemini 3.5 Flash when Gemini 3.6 Flash has the same standard input price and a lower output price. Existing applications, however, should still test migration compatibility instead of changing the model ID without review.

Developers who are not ready to migrate can continue to test Gemini 3.5 Flash on GPT Proto before comparing its real output quality, latency, and token consumption with Gemini 3.6 Flash.

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite Pricing

The following prices are Google’s paid Gemini API rates per one million tokens at launch.

*Gemini 3.1 Flash-Lite charges a higher input rate for audio. Refer to the current Gemini API pricing page before deploying a production workload.

A Simple Cost Example

Suppose a workload consumes 100 million input tokens and generates 20 million output tokens.

Ignoring caching, grounding, and storage fees, the standard Google API cost would be:

• Gemini 3.6 Flash: $150 input + $150 output = $300
• Gemini 3.5 Flash: $150 input + $180 output = $330
• Gemini 3.5 Flash-Lite: $30 input + $50 output = $80

This illustrates the large sticker-price difference between Flash and Flash-Lite. However, fixed token calculations do not tell the entire story.

A cheaper model can become expensive if it requires more retries, produces longer reasoning traces, or fails to complete a task. Similarly, a model with a higher token rate may have a lower cost per completed task if it uses fewer turns and tools.

That is why Gemini 3.6 Flash’s token efficiency matters. If it produces approximately 17% fewer output tokens while also charging 16.7% less for each output token, the savings on the output portion of some workloads can be substantially greater than the price-table difference alone suggests.

Gemini 3.5 Flash-Lite presents the opposite lesson. It is the least expensive model in the 3.5 family, but it is not cheaper than 3.1 Flash-Lite for every token type. Its text, image, and video input price increased from $0.25 to $0.30 per million tokens, while output increased from $1.50 to $2.50.

Developers are paying more than they did for 3.1 Flash-Lite, but they are receiving much better reasoning, agentic performance, and throughput.

GPT Proto Pricing: Access Both Models at 40% Off

Both models are also available through GPT Proto at 40% below Google’s listed standard token rates.

At these rates, the same workload of 100 million input tokens and 20 million output tokens would cost approximately:

• Gemini 3.6 Flash on GPT Proto: $90 input + $90 output = $180
• Gemini 3.5 Flash-Lite on GPT Proto: $18 input + $30 output = $48

The same workload would cost approximately $300 and $80 respectively at Google’s standard listed rates.

GPT Proto uses pay-as-you-go billing, so developers do not need a separate subscription for each model. The same API key and account balance can also be used across GPT Proto’s collection of 200+ text, image, video, and audio models.

You can check the current rates on the Gemini 3.6 Flash API page and Gemini 3.5 Flash-Lite API page.

Which Gemini Flash Model Should You Choose?

The best model depends on whether your bottleneck is task complexity or execution volume.

The More Interesting Answer: Use Both

Imagine an e-commerce platform that needs to extract product information from thousands of listings, PDFs, images, and supplier documents.

Gemini 3.6 Flash could:

1. Inspect several representative documents.
2. Design the extraction schema.
3. Decide which sources are reliable.
4. Identify ambiguous or conflicting product information.
5. Review exceptions that require deeper reasoning.

Gemini 3.5 Flash-Lite could then run in parallel to:

1. Read every product document.
2. Extract brand, material, size, price, and availability.
3. Translate product descriptions.
4. Return structured JSON.
5. Flag unusual records for the main agent.

This architecture avoids paying 3.6 Flash rates for every repetitive extraction while still using the stronger model where its reasoning matters.

Gemini 3.6 Flash vs Kimi K3

Gemini 3.6 Flash and Kimi K3 were released within days of each other, but they optimize for different things.

Kimi K3 is Moonshot AI’s 2.8-trillion-parameter flagship model for long-horizon coding, reasoning, and end-to-end knowledge work. Gemini 3.6 Flash emphasizes speed, token efficiency, multimodal workflows, and integration with Google’s tools.

The following figures come from the same Artificial Analysis comparison, making them more useful than comparing unrelated vendor benchmark tables.

Kimi K3 has the stronger overall Intelligence Index score. It is a better candidate when maximum reasoning quality and sustained knowledge work matter more than response speed.

Gemini 3.6 Flash is approximately twice as cheap by listed input and output token prices, and its observed generation speed is several times higher. That makes it more attractive for interactive tools, coding loops, document analysis, and applications serving many simultaneous users.

There is also an openness difference, although it requires a date-sensitive qualification. Moonshot described Kimi K3 as an open-source model, but its official launch documentation said the complete weights would be released by July 27, 2026. They were not yet fully available at this article’s July 23 update.

The practical verdict is:

• Choose Gemini 3.6 Flash for speed, lower API cost, multimodal input, and Google-native tools.
• Choose Kimi K3 for stronger general intelligence, long-running knowledge work, and future self-hosting possibilities.
• Test both on the actual workflow before assuming a higher benchmark score will produce a lower cost per successful task.

You can also try Kimi K3 on GPT Proto to compare its behavior with other current models.

Where Can You Access Gemini 3.6 Flash and Gemini 3.5 Flash-Lite?

Google provides direct access to the two models through Google AI Studio, the Gemini API, the Gemini app, and its enterprise products. Gemini 3.6 Flash is also available through Google Antigravity.

Developers who want one key for multiple AI providers can now access both models through GPT Proto:

• Gemini 3.6 Flash API on GPT Proto
• Gemini 3.5 Flash-Lite API on GPT Proto

GPT Proto provides an OpenAI-compatible API interface. Instead of maintaining separate Google, Kimi, OpenAI, DeepSeek, and other provider accounts, developers can use one API key, one balance, and the same base URL across 200+ models.

The GPT Proto model strings are:

gemini-3.6-flash
gemini-3.5-flash-lite

This also makes A/B testing easier. An application can route complex requests to Gemini 3.6 Flash and high-volume extraction or classification jobs to Gemini 3.5 Flash-Lite without rebuilding its API integration.

How to Use Gemini 3.6 Flash and Gemini 3.5 Flash-Lite API on GPT Proto

GPT Proto supports an OpenAI-compatible Chat Completions endpoint, so developers who already use the OpenAI SDK can access the two Gemini models by changing the API key, base URL, and model name.

Step 1: Create a GPT Proto API Key

Create or sign in to your GPT Proto account, add usage credits, and open the API Keys section in the dashboard.

Generate a new API key and store it securely. Do not paste the key directly into public repositories, frontend code, or shared documents.

For macOS or Linux, save it as an environment variable:

export GPTPROTO_API_KEY="sk-your-gptproto-api-key"

Step 2: Install the OpenAI Python SDK

Install or update the OpenAI SDK:

python -m pip install openai

GPT Proto uses the following OpenAI-compatible base URL:

https://gptproto.com/v1

Step 3: Call the Gemini 3.6 Flash API

Use Gemini 3.6 Flash when the request involves coding, planning, multimodal analysis, knowledge work, or a complex sequence of decisions.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GPTPROTO_API_KEY"],
    base_url="https://gptproto.com/v1",
)

response = client.chat.completions.create(
    model="gemini-3.6-flash",
    messages=[
        {
            "role": "system",
            "content": (
                "You are a careful software engineering assistant. "
                "Inspect the problem before proposing changes and explain "
                "how each recommendation should be verified."
            ),
        },
        {
            "role": "user",
            "content": (
                "Review this API migration plan and identify compatibility, "
                "security, and performance risks."
            ),
        },
    ],
    max_completion_tokens=2048,
)

print(response.choices[0].message.content)

This request is sent to GPT Proto’s unified endpoint while using gemini-3.6-flash as the model ID.

Step 4: Switch to Gemini 3.5 Flash-Lite

To use Flash-Lite, keep the same client, API key, and base URL. Only change the model string:

response = client.chat.completions.create(
    model="gemini-3.5-flash-lite",
    messages=[
        {
            "role": "user",
            "content": (
                "Extract the merchant, invoice date, currency, subtotal, tax, "
                "and total from the following invoice text. Return valid JSON."
            ),
        }
    ],
    max_completion_tokens=1024,
)

print(response.choices[0].message.content)

Gemini 3.5 Flash-Lite is the better option for large batches of extraction, translation, classification, routing, and other clearly defined tasks.

Step 5: Route Tasks Between the Two Models

A production application does not have to send every request to the same model. A simple routing rule can use:

• gemini-3.6-flash for planning, coding, exception handling, and difficult multimodal analysis.
• gemini-3.5-flash-lite for repetitive execution, structured extraction, translation, and high-volume processing.

Because both models use the same GPT Proto endpoint and account balance, switching between them only requires changing the model ID.

Avoid adding deprecated Gemini sampling parameters such as temperature, top_p, and top_k to new integrations. Google has deprecated these parameters for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite.

Migration Notes: Do Not Treat It as a Model-ID-Only Upgrade

Gemini 3.6 Flash may look like an obvious replacement for Gemini 3.5 Flash, but Google introduced API behavior changes that can affect existing applications.

Starting with Gemini 3.6 Flash and Gemini 3.5 Flash-Lite:

• temperature is deprecated.
• top_p is deprecated.
• top_k is deprecated.
• Prefilled model turns are no longer supported.
• A request whose last non-empty turn is a model message can return an HTTP 400 error in future implementations.

Google currently says the deprecated sampling parameters are ignored, but future model generations may reject them. Developers should remove them rather than relying on the API to continue ignoring the values.

Before migrating production traffic, test:

• Structured JSON output
• Function and tool calls
• Multi-turn message history
• Prompts that previously depended on temperature
• Long-document accuracy
• Thinking-level behavior
• Token use and cost per completed task
• Retry and timeout handling

This is particularly important for agentic applications, where a small change in tool selection or reasoning length can significantly affect total cost.

Limitations to Know Before Switching

Both new models are capable, but they still have important limitations.

They Only Generate Text

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite can analyze images, videos, audio, and PDFs, but they do not directly generate media. Image, video, speech, and music generation require separate models.

A 1M Context Window Does Not Guarantee Perfect Recall

The context limit tells you how much content the API can accept, not whether every fact will receive equal attention. Important instructions and evidence should still be clearly structured, especially in very long prompts.

Thinking Tokens Count Toward Output Billing

Increasing the thinking level can improve difficult tasks, but it also increases output consumption and latency. Flash-Lite at a high thinking level may behave very differently from its minimal default.

GA Does Not Eliminate Hallucinations

Both official model cards list hallucination, occasional slowness, and timeouts as known limitations. High-stakes outputs still require grounding, validation, or human review.

Computer Use Documentation Is Currently Inconsistent

Google’s launch guide states that the new models support its built-in tool suite, including Computer Use. However, the individual Gemini 3.5 Flash-Lite model page currently lists Computer Use as unsupported, while the launch materials describe it as available.

Developers planning browser or interface automation with Flash-Lite should verify the latest API capability documentation before production deployment.

Final Verdict

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are not redundant releases.

Gemini 3.6 Flash is the better default when a task requires coding, planning, complex reasoning, multimodal interpretation, or multiple tool calls. Its main advantage over Gemini 3.5 Flash is not a dramatic increase in general intelligence, but a better combination of speed, token efficiency, coding reliability, and output pricing.

Gemini 3.5 Flash-Lite is the better choice when the task is clearly defined and must be repeated thousands or millions of times. It costs more than the previous Flash-Lite generation in several pricing categories, but it delivers a significant improvement in reasoning, tool reliability, and throughput.

For many production systems, the best decision will be to use both: Gemini 3.6 Flash as the planner and Gemini 3.5 Flash-Lite as the scalable execution layer.

Both models are now available on GPT Proto at 40% off Google’s standard listed rates. Developers can use the Gemini 3.6 Flash API for complex coding and agentic workflows, switch to the Gemini 3.5 Flash-Lite API for high-volume execution, or browse the full GPT Proto AI model collection to compare more than 200 models through one API platform.