Gemini 3.7 Flash Released Suddenly! Catch Up on the Overview and Features Quickly

Last Updated: 2026/8/14

"Gemini 3.6 Flash just came out, and now the next model?"
Google announced the new model Gemini 3.7 Flash on its official English blog on August 13, 2026, and on its official Japanese blog on August 14.
What is surprising is the release interval.
It has been only about three weeks since the release of Gemini 3.6 Flash. Google positions 3.7 Flash as its "most capable flagship model to date" for coding and agent tasks.⟦S1⟧ ⟦S2⟧
Moreover, it is not just a performance boost.
The paid API price until December 31, 2026, is half the initial price of Gemini 3.6 Flash. It is offered at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens.⟦S1⟧ ⟦S5⟧
To describe this update in one phrase,

is what it is.

In this article, we will organize the points regarding Gemini 3.7 Flash, focusing on the following:

1. What kind of model is Gemini 3.7 Flash?

2. How much have coding and web development capabilities improved?

3. How have pricing and cost-performance changed?

4. How do you use it via API or Gemini Spark?

5. What are the constraints and points to note before migrating?

*Created based on official information as of August 14, 2026.

*The created article content was input into Gamma to automatically generate slides. Please take a look here if you find slides easier to read.

Summary of Gemini 3.7 Flash Key Points

First, let's summarize the key points of this update.

• Gemini 3.7 Flash is a generally available (GA) model and is recommended for production use.

• It was released just about three weeks after Gemini 3.6 Flash.

• Performance has improved in coding, multi-step agent processing, web development, and knowledge work.

• It supports a 1 million token input context and up to 65,536 output tokens.

• You can select the thinking level from three stages: low, medium, and high

• Paid API pricing through the end of 2026 is $0.75 for input and $3.75 for output

• Individuals can use it via Gemini Spark, developers via AI Studio or the API, and enterprises via the Gemini Enterprise environment

⟦S1⟧ ⟦S3⟧ ⟦S4⟧
Gemini 3.7 Flash is a step beyond the traditional lightweight model image of being 'cheap and fast but bad at difficult tasks'.

1. First of all, what is Gemini 3.7 Flash?

Gemini 3.7 Flash is a native multimodal reasoning model belonging to the Gemini 3 series.
Google cites complex code generation, multi-step agent processing, design-faithful web development, and enterprise business workflows as its primary use cases.
It is already generally available (GA) in the API documentation and is presented as a model that can be used in production environments.⟦S3⟧
Basic specifications of Gemini 3.7 Flash

The ability to handle long inputs of 1 million tokens while passing not just text but also images, videos, audio, and PDFs to a single model is a strength shared with the previous Flash series.
In this 3.7 Flash, the ability to perform complex tasks involving code modification and tool usage while referencing that long information has been significantly enhanced.

2. The biggest evolution is 'making it harder to stop complex tasks midway'

What to pay attention to in Gemini 3.7 Flash is not just the accuracy of a single response.
Google explains that 3.7 Flash has improved in handling issues that occur during work, confirming intent as needed, multi-step planning, tool calling, and instruction following.⟦S1⟧
Agent-type work usually proceeds in the following flow:

With previous models, there were times when they would repeat the same mistakes, think longer than necessary, or choose the wrong tools.
3.7 Flash has been improved to stabilize such 'agent loops' and reduce failed processing or manual retries.⟦S3⟧
However, the official explanation does not guarantee that rework in your company's environment will definitely decrease.
What you really need to check is not the benchmark ranking, but

• Whether the task was completed to the end

• How many times it retried

• How many times it called tools

• How much human correction was required

• How much it cost per completed job

These are the practical results.

3. Where did it improve in benchmarks?

The main comparison results released by Google are as follows.

Particularly notable are DeepSWE, which evaluates long-term software development, AutomationBench, which evaluates enterprise workflows, and OSWorld-2.0, which includes screen operation.
Not just single-shot code generation, but

• Examine the repository

• Hypothesize the cause

• Modify multiple files

• Execute commands

• Verify the results

The difference appears in evaluations related to long tasks such as these.
However, the Gemini 3.6 Flash score in DeepSWE is 49.0% in Google's announcement blog and 48.6% in the model card. Since the timing of the evaluation and the aggregation methods in the materials are not clearly stated, it is safer to include the source when citing these figures.⟦S1⟧ ⟦S6⟧

4. Web development has improved not only in "appearance" but also in "completeness"

In web development, reproducibility has improved when providing screenshots, reference images, design systems, etc.
According to Google, Gemini 3.7 Flash is now capable of generating functional applications with practical layouts using fewer prompts.
The WebDev Arena Elo score increased from 1538 for Gemini 3.6 Flash to 1588 for 3.7 Flash.⟦S1⟧
What is noteworthy is not just outputting HTML that resembles a reference image.

• Buttons work

• Necessary screen transitions are prepared

• Components are reusable

• Specified design rules are reflected

• Differences between existing code and design proposals can be checked

These are cited as improvements in the ability to approach a finished product.⟦S3⟧
It seems worth trying for tasks such as creating front-end code from designer mockups or auditing design inconsistencies in existing screens.

5. Stronger in PDF and corporate knowledge work

This evolution is not limited to the development field.
In GDP.pdf, which evaluates the understanding of complex PDF documents, it improved from 22.0% for Gemini 3.6 Flash to 34.0% for 3.7 Flash.
Also, in AutomationBench, which assumes actual corporate workflows, it grew from 17.0% to 30.4%.⟦S1⟧ ⟦S6⟧
Easily imaginable use cases include the following:

• Extracting important figures and points of discussion from annual reports

• Cross-referencing contracts with internal regulations

• Creating reports across multiple PDFs

• Analyze tables, graphs, and body text all at once

• Convert research results into web content

• Complete routine tasks using multiple business tools

However, AutomationBench is listed as a "Private set" on the model card.
Since not all test data is public, it is best to avoid judging its suitability for your business based solely on scores.⟦S6⟧

6. What can 1M tokens and multimodality be used for?

Gemini 3.7 Flash supports inputs of up to 1,048,576 tokens and outputs of 65,536 tokens.
Because you can input images, videos, audio, and PDFs in addition to text, you can analyze materials in different formats together.⟦S4⟧
For example, the following use cases are possible.

• Load a large codebase and specifications simultaneously

• Organize meeting audio, minutes, and related materials across the board

• Analyze the body text and graphs of financial statements together

• Find specific scenes or points of discussion from long videos

• Compare UI images with existing code to create correction proposals

• Extract differences in terms from multiple contracts

However, being able to "input 1 million tokens" is not the same as "being able to understand the entire 1 million tokens with the same accuracy at all times."
When handling long documents, it is more stable to clearly specify document breaks, reference priorities, required citation formats, and response formats.
For important decisions, it is necessary to design the system to output not only the conclusion but also the source file name, page, and relevant section.

7. How has the pricing changed?

As of August 14, 2026, the paid API pricing for Gemini 3.7 Flash is as follows.
All prices are per 1 million tokens, and output prices include thinking tokens.⟦S5⟧

Standard API

Batch API / Flex inference

A free tier is also available.
Also, this introductory pricing applies not only to Gemini 3.7 Flash but also to Gemini 3.6 Flash. From January 1, 2027, both models will be priced at $1.50 for input and $7.50 for output for the Standard API.⟦S1⟧ ⟦S5⟧

Do not judge "cheap" based on unit price alone
In agent-based processing, the number of tokens used for a single job may increase due to thinking tokens, tool calls, retries, and long outputs.
When introducing it, it is easier to understand if you think using the following formula.

Even with a cheap model, if you fail repeatedly, the total cost will be high.
Conversely, even if the consumption per task increases, if the work is completed in one go and human correction time is reduced, the total cost may decrease.

8. How to use it

Gemini 3.7 Flash can be used from multiple environments depending on the type of user.

Developers

• Google AI Studio

• Gemini API

• Android Studio

• Google Antigravity

Corporate

• Gemini Enterprise Agent Platform

• Gemini Enterprise App

Individual

• Gemini Spark for Google AI Pro/Ultra users

Gemini Spark is an always-on personal AI agent provided to Google AI Pro/Ultra users in over 160 countries.
It is introduced for use in coordinating with Google Workspace apps, organizing files, drafting emails, and updating status documents.⟦S1⟧ ⟦S2⟧
Note that availability for individual users does not mean that everyone can select it directly from the standard model selection screen. The official announcement explains that it is available to Google AI Pro/Ultra users through Gemini Spark.

Using it from the Gemini API
The model ID is as follows.

gemini-3.7-flash

A basic Python example using the Interactions API is as follows.⟦S3⟧

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-3.7-flash",
    input=(
        "このコードベースを確認し、不具合の原因を調査してください。"
        "影響範囲を整理したうえで、安全な修正計画を提案してください。"
    ),
    generation_config={
        "thinking_level": "medium"
    },
)

print(interaction.output_text)

Differentiating by thinking level

The default is medium.
With high, you can increase thinking and tool usage, but token consumption, costs, and processing time may also increase.⟦S3⟧
Also, Gemini 3.7 Flash is now the new default model for Antigravity agents used in Gemini Managed Agents and the Google Antigravity SDK.⟦S3⟧

9. Four points to check before implementation

Point 1: Benchmarks do not guarantee practical quality
Published scores are a powerful tool for comparing models.
However, tools, thinking levels, trial conditions, and evaluation data differ for each evaluation. Also, self-measurements by Google and non-public evaluation sets are included.⟦S6⟧
A/B testing using your own code, documents, tools, and termination conditions is necessary.

Note 2: It did not outperform 3.6 Flash in all evaluations
In the model card, the tool-free evaluation on CharXiv, which integrates information from complex graphs, was 84.5% for 3.7 Flash compared to 85.2% for 3.6 Flash.
Even with tools, it is 88.7% for 3.7 Flash compared to 89.4% for 3.6 Flash.⟦S6⟧
Just because it is a new model does not mean quality will automatically improve for all inputs or use cases.

Note 3: Computer Use is in preview
Gemini 3.7 Flash supports Computer Use, but it is currently in preview.⟦S4⟧
For automation involving screen operations, at least the following measures are required.

• Run in an isolated environment for verification

• Restrict operable sites and functions

• Require human approval for sending emails, making purchases, and deletions

• Log all operations

• Set maximum operation counts and timeouts

• Minimize access permissions to production data

Note 4: Hallucinations and timeouts remain
The model card explains that hallucinations may occur as a general limitation of foundation models.
Also, processing may become slow or timeouts may occur.
The knowledge cutoff is generally March 2026, but information in some fields may be limited to what was available as of January 2025.⟦S6⟧
When dealing with the latest information, you must combine it with Google Search grounding, reliable internal search, RAG, and verification against primary sources.

10. How have safety measures changed?

In Gemini 3.7 Flash, safeguards have been updated for the CBRN domain—chemical, biological, radiological, and nuclear materials—as well as for offensive cyber use.⟦S1⟧ ⟦S6⟧
The model card indicates that it has not reached the critical capability thresholds that are tracked.
On the other hand, the fact that safety measures have been updated does not mean it is 'safe for any use.'
When using it in a corporate environment, do not rely solely on model-side controls; you must combine them with:

• Access management

• Input data classification

• Approval for external operations

• Audit logs

• Output verification

• Incident shutdown procedures

as well.

11. Tasks suited for Gemini 3.7 Flash and tasks to handle with caution

Gemini 3.7 Flash is not a universal model to be applied unconditionally to all processes.
It provides the most value in tasks that do not end with a single response, but rather involve repeating investigation, planning, tool operation, correction, and verification.

Suitable tasks

• Investigating large-scale codebases

• Code modifications spanning multiple files

• Root cause analysis of bugs

• AI agents using multiple tools

• Web implementation from screenshots

• Auditing differences between designs and existing screens

• Document analysis across PDFs, tables, and graphs

• Development based on long specifications

• Workflow automation using Google Workspace

• Multi-step report creation

Tasks where other low-cost models are also worth considering

• Simple classification

• Routine summarization

• Large-scale data extraction

• Short translations

• Conversion to structured JSON

• Processing that generates many ideas in parallel

• Simple subtasks within agents

Tasks requiring caution

• Payments or purchases without human approval

• Irrecoverable file deletion

• Unconditional code deployment to production environments

• Unsupervised screen operations involving sensitive information

• Final decision-making in legal, medical, or financial fields

• Research articles published without source verification

• Real-time voice conversation

• Image generation by the model alone

I do not believe the criterion for choosing a model is whether it is the 'newest.'
I believe it is necessary to make decisions based on the impact of failure, processing volume, required accuracy, acceptable latency, and the extent to which humans can verify the results.

12. Where to start testing

If you are already using Gemini 3.6 Flash, I recommend starting with small A/B tests rather than switching everything at once.

Step 1: Choose practical tasks
Select several tasks that typically consume a lot of time or require frequent rework.
For coding, tasks like the following are more suitable than simple function generation:

• Investigating the cause of bugs

• Modifying multiple files

• Adding tests

• Refactoring while maintaining existing design

• Implementing screens from design mocks

Step 2: Standardize conditions
Set the same inputs, same tools, and same completion criteria for both 3.6 Flash and 3.7 Flash.
Unless the reasoning level is also aligned, it will not be a fair comparison.

Step 3: Record the results
Measure the following items:

• Task completion rate

• Percentage of initial outputs accepted as-is

• Input/Output token count

• Number of tool calls

• Number of retries

• Processing time

• Unintended changes

• Time spent on human corrections

Step 4: Review failure details
I think there is a lot to learn from checking not just the average score, but also how it failed. For example,

• Used a non-existent API

• Modified files outside the scope

• Completed without passing tests

• Repeated the same operation

• Provided baseless numerical values

Failures like these directly impact operational design.

Step 5: Calculate using 2027 pricing
Estimate total costs using not only the introductory price but also the regular price effective from January 1, 2027. Even if it is profitable at year-end rates, your decision might change once the price doubles.

Summary: Gemini 3.7 Flash is a model for "running difficult tasks at Flash prices"

If you view Gemini 3.7 Flash simply as "a model with higher numbers than Gemini 3.6 Flash," you will miss the essence of this update. To summarize the key points:

• Released just about 3 weeks after Gemini 3.6 Flash

• A flagship model focused on coding and agent use cases

• Improved performance in FrontierCode, DeepSWE, WebDev Arena, etc.

• Scores have improved in PDF understanding and enterprise workflows

• Supports 1M token input and 64k token output

• Paid API pricing through the end of 2026 is $0.75 for input and $3.75 for output

• Available for individuals via Spark, developers via AI Studio and API, and enterprises via Enterprise environments

• Computer Use is in preview, with limitations such as hallucinations still remaining

What is important about this evolution is not just competing for the highest scores.

I believe the Flash series has moved in a direction directly linked to practical work.
Gemini 3.7 Flash is neither just a "superior version of 3.6 Flash" nor an "inexpensive all-purpose model."

It is a flagship model for maintaining long context and handling difficult tasks on a daily basis while using necessary tools.

If you recognize it that way, I think the meaning of this release will be easier to understand.
First, it is helpful to choose one high-difficulty task that currently requires the most manual labor and compare it under the same conditions as Gemini 3.6 Flash.
I believe you will only see its true value for your company by measuring not just the impression of the response, but also the task completion rate, number of retries, human correction time, and total cost including standard pricing.

References/Sources

• ⟦S1⟧ Google Japan Blog "Announcing Gemini 3.7 Flash" https://blog.google/intl/ja-jp/company-news/technology/gemini-37-flash/

• ⟦S2⟧ Google Blog "Introducing Gemini 3.7 Flash" https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/

• ⟦S3⟧ Google AI for Developers "What's new in Gemini 3.7 Flash" https://ai.google.dev/gemini-api/docs/latest-model

• ⟦S4⟧ Google AI for Developers "Gemini 3.7 Flash" Model Specifications https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash

• [S5] Google AI for Developers "Gemini Developer API pricing" https://ai.google.dev/gemini-api/docs/pricing

• [S6] Google DeepMind "Gemini 3.7 Flash Model Card" https://deepmind.google/models/model-cards/gemini-3-7-flash/