Google Announces New Model "Gemini 3.7 Flash" in Just Three Weeks — Coding Performance Up 65%, Half Price Through Year-End
機械翻訳 / Machine-translated
A
AI Friends
9時間前・15分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
On August 13, 2026, Google announced a new AI model called "Gemini 3.7 Flash." It arrives just three weeks after the previous model, 3.6 Flash, and comes with a surprising strategy: half-price through the end of the year.
What you'll learn in this article:
Gemini 3.7 Flash is Google's latest generative AI model. The previous version, 3.6 Flash, had only just launched in July 2026, yet the next version was announced a mere three weeks later.
There is a reason for this unusually fast pace: competition in the fields of coding (programming) and agents (systems where AI performs tasks autonomously) is intensifying. Google is aiming to win developer support by rapidly iterating on improvements.
The defining characteristics of Gemini 3.7 Flash are the combination of "coding specialization" and "low pricing." Traditional AI models were good at writing text and answering questions, but their ability to write program code still left much to be desired. This model, however, aims to break through that barrier in a major way.
Gemini 3.7 Flash is priced at $0.75 for input and $3.75 for output per million tokens (a unit of text length). This is exactly half the price of the previous model, 3.6 Flash.
However, this is an "introductory price" — a time-limited special rate. Usage through December 31, 2026 is half price, but starting January 1, 2027, prices will revert to standard rates ($1.50 for input, $7.50 for output). In other words, the more you use it before year-end, the better the deal.
This pricing strategy runs counter to the current trend in the AI industry. That's because in early August 2026, China's DeepSeek announced it would be significantly raising its API fees in the near future. DeepSeek had attracted a large number of developers with its ultra-low pricing, but soaring demand led to ballooning costs, prompting the price hike.
Google, on the other hand, is going on the offensive with a price cut. By announcing a half-price offer right as DeepSeek raises its prices, Google appears to be aiming to draw developers over to Gemini. Fierce price competition is currently playing out across the AI industry, and which model wins developer loyalty may well determine the outcome.
Gemini 3.7 Flash has significantly outperformed its predecessor on coding benchmarks (performance tests). Here are some examples:
These figures indicate how accurately the AI can write programs, find and fix bugs, and handle automation tasks. Compared to 3.6 Flash, 3.7 Flash shows improvements of more than 10–15 percentage points across every test.
Particularly noteworthy is the DeepSWE v1.1 score. This test measures "software engineering autonomy" — how well an AI can perform like a human engineer. A score of 65.3% means the AI is approaching a level where it can handle roughly two-thirds of what a human would do.
The model is also capable of generating well-structured layouts and apps in web development with fewer prompts (instructions). In other words, even without detailed instructions from a developer, the AI can understand context and write appropriate code.
The arrival of Gemini 3.7 Flash brings significant benefits for businesses and developers. First, on the cost side: by taking advantage of the half-price introductory rate, organizations can substantially reduce expenses for AI-powered development and automation.
For example, a company that performs one million tokens' worth of code reviews and auto-generation per day would pay roughly $135 (approximately ¥20,000) per month combined for input and output. At standard pricing, that rises to around $270 (approximately ¥40,000) — so introducing the model before year-end means they can trial it at half the cost.
Next comes the productivity boost. With enhanced coding capabilities, developers can delegate routine, repetitive work to the AI and focus on more creative tasks. For instance, if the AI handles automatic test code generation, bug fix suggestions, and documentation writing, development speed can increase dramatically.
On top of that, agent capabilities have also been strengthened. An agent is a system where an AI independently makes decisions and carries out multiple tasks in sequence. For example, a series of steps like "retrieve data from a database, analyze it, and generate a report" can be executed automatically without requiring a human to give instructions at every step. This is expected to push business automation even further.
Gemini 3.7 Flash has added a feature that lets you choose the "inference load" — essentially, how deeply the AI thinks when generating an answer, i.e., how much computational processing it applies.
There are three options: "low," "medium," and "high." For simple questions or code completion, using "low" is fast and inexpensive; for complex problem-solving, "high" delivers better accuracy. By matching the setting to the task, users can balance cost and performance accordingly.
This is particularly convenient for businesses processing large volumes of tasks. For example, "low" could be used for simple log analysis while "high" is reserved for reviewing important contracts.
The release of Gemini 3.7 Flash signals that competition in the AI industry has entered a new phase. With major models such as OpenAI's GPT, Anthropic's Claude, and DeepSeek all vying fiercely against one another, Google is trying to differentiate itself through "speed" and "price."
A higher-tier model, "Gemini 3.5 Pro," is also expected in the future, though no specific timeline has been announced. That said, given that the update from 3.6 to 3.7 took only three weeks, the next version could potentially arrive quite soon as well.
Across the AI industry as a whole, "practical usefulness" is increasingly emphasized. It's no longer just about pushing benchmark numbers higher — what matters is how much value the model provides in real development environments and business operations. Gemini 3.7 Flash can be said to have taken a step ahead on that front.
You can start using Gemini 3.7 Flash immediately in the following environments:
For first-time users, Google AI Studio is recommended. Since it runs in a browser, no installation is required and it's easy to try out. Simply entering a basic prompt lets you experience the power of code generation and debugging firsthand.
If you plan to use the API, you'll need to create a Google Cloud account and generate an API key. The official documentation includes detailed steps, so even beginners can get started with confidence.
The AI development race looks set to intensify further. In that landscape, Gemini 3.7 Flash — armed with practicality and cost-effectiveness — is poised to become a strong ally for developers and businesses alike.
This article is a cross-post from AI Friends.
