What has changed in Gemini 3.7 Flash?
What has changed in Gemini 3.7 Flash?
—A major update from the previous generation Gemini 3.6 Flash
In August 2026, Google introduced the new "Gemini 3.7 Flash."
The previous generation, Gemini 3.6 Flash, was released on July 21, 2026.
In other words, it was updated to 3.7 Flash in a very short period of just a few weeks.
So, what has changed from 3.6 Flash to Gemini 3.7 Flash?
In conclusion,
The key point is that it has evolved from a "fast AI" to an "AI that can carry out more complex tasks to completion while maintaining its speed."
—
1. The biggest evolution is in "coding"
Gemini 3.6 Flash was also a model with significantly enhanced coding performance.
In 3.7 Flash, this has been further improved in areas such as:
* Code generation
* Bug fixing
* Web development
* Multi-file corrections
* Program comprehension
* Debugging
* Long development tasks
It is particularly important that it has become stronger not just at "writing code," but at the entire process of reading code, discovering problems, fixing them, and searching for the next issue.
This also highlights the difference between a standard question-answering AI and an AI coding agent.
—
2. Enhanced multi-step reasoning
3.7 Flash has improved support for tasks that do not end with a single response.
For example, instead of just:
"Fix this HTML"
it is designed with a greater awareness of multi-stage tasks such as:
1. Reviewing the entire HTML
2. Checking the JavaScript structure
3. Identifying the cause of the bug
4. Fixing the relevant code
5. Checking for impacts elsewhere
6. Verifying the state after the fix
Google positions 3.7 Flash as a model specifically for software development and AI agent workflows. (Reuters)
—
3. "Instruction following" is also a key update
When you give complex instructions to an AI, the conditions often pile up, such as:
"Fix A"
"However, do not change B"
"Keep the design as is"
"Check the operation at the end"
Previous Flash models sometimes forgot parts of these instructions when given multiple conditions.
In 3.7 Flash, the ability to work while maintaining multiple conditions has been enhanced.
Therefore, it is better suited for development tasks like "fix only this part and do not change anything else."
—
4. Evolution for AI agents
Another major feature of 3.7 Flash is its evolution from an "AI that answers" to an "AI that works."
For example, with an AI agent, after receiving an instruction like "Please fix this website," it needs to perform tasks such as:
Investigation
↓
Planning
↓
Code modification
↓
Verification
↓
Additional fixes
3.7 Flash is a model strongly designed with such multi-stage agent workflows in mind.
In other words, it is moving closer to a work flow of:
Instruction → Planning → Execution → Verification → Correction
rather than just: Question → Answer
—
5. Even better compatibility with Web development
Web development is of particular note. In web apps using HTML, CSS, and JavaScript, many elements are interrelated simultaneously, such as:
* HTML structure
* CSS layout
* JavaScript processing
* Button events
* Databases
* Screen transitions
* Scrolling
* Responsive design
3.7 Flash has undergone performance improvements with a greater awareness of such complex development tasks.
Therefore, it can be said that its usability as an "AI that fixes web apps" has increased compared to just being an "AI that generates code."
—
6. Ability to handle long tasks
With large programs, simply looking at "this part" may not solve the problem.
For example, if a single HTML file contains:
* Thousands of lines of HTML
* Massive amounts of CSS
* JavaScript
* Problem data
* Settings
* Various functions
In such code, fixing one part locally can sometimes break another.
3.7 Flash has been improved with a greater awareness of such long code and multi-step tasks.
—
7. The "Flash" philosophy of increasing performance while maintaining speed
This is the interesting part of 3.7 Flash.
It is not a simple direction of "if you make it high-performance, it becomes slow."
The Flash series has always prioritized the balance of performance, speed, and cost.
With 3.6 Flash, Google also emphasized high token efficiency, low latency, and reliability. It is also noted that 3.6 Flash reduced output token usage by 17% compared to 3.5 Flash. (blog.google)
3.7 Flash further develops that direction.
—
8. Significant gains in benchmarks
3.7 Flash reports significant gains from 3.6 Flash, especially in evaluations related to coding, web development, and agents.
For example, in published comparisons:
Field Gemini 3.6 Flash Gemini 3.7 Flash
FrontierCode 1.1 Main 34.4% 43.6%
DeepSWE v1.1 49.0% 65.3%
WebDev Arena 1538 Elo 1588 Elo
GDP.pdf 22.0% 34.0%
AutomationBench 17.0% 30.4%
Such significant improvements have been reported.
The gain in AutomationBench, in particular, shows that the focus is on the "ability to actually carry out work" rather than simple question-answering.
※ Since benchmark results change depending on measurement conditions, overall performance cannot be judged by these numbers alone.
—
9. In a word, the difference from 3.6 Flash
Gemini 3.6 Flash: "A Flash model that handles work quickly and efficiently"
↓
Gemini 3.7 Flash: "A Flash model that advances complex development tasks longer and more accurately while maintaining speed"
That is the image.
—
10. Who benefits the most
3.7 Flash is particularly suitable for:
* People who program
* People who fix HTML/CSS/JavaScript
* People who build web apps
* People who use AI coding
* People who use AI agents
* People who handle large codebases
* People who want to assign tasks with multiple conditions
On the other hand, if you just want "simple questions answered," there are cases where you might not feel a dramatic difference from 3.6 Flash.
This update can be said to be one where those who "have AI do difficult work" will benefit the most.
—
Summary
Gemini 3.7 Flash is not simply a "fast version of 3.6."
The major direction is an evolution from:
3.6 Flash (Fast, low-cost, high-efficiency)
↓
3.7 Flash (Coding enhancement + multi-step reasoning + instruction following + web development + AI agents + complex task execution capability)
What is particularly noteworthy is that it has enhanced not only the "ability to provide answers" but also the "ability to carry work through to completion."
I think this is the biggest point of Gemini 3.7 Flash.
And for those who want to use AI not just as a "tool to ask questions," but as a "development partner that writes, fixes, and checks code together," 3.7 Flash is a model that deserves significant attention.
※ Since Gemini 3.7 Flash was just announced on August 13, 2026, usage environments and evaluation results may be updated further in the future. (Reuters)