Just three weeks after Gemini 3.6 Flash, Google is already picking up the pace with Gemini 3.7 Flash, a new model designed to boost performance in programming, automation and workflows involving multiple artificial intelligence agents. This particularly fast cadence illustrates the intensifying competition among the major labs, which can no longer wait several months between two model generations to incorporate technical progress and developer feedback.¹
With Gemini 3.7 Flash, however, Google is not just trying to gain a few extra points on benchmarks. The company wants to offer a model powerful enough to handle complex professional tasks, yet fast and affordable enough to be used at scale in applications, AI agents and development environments. This strategy comes with a particularly aggressive selling point: a launch price well below that of its predecessor, increasing the pressure on OpenAI, Anthropic and the other players locked in the price-performance battle.²
Gemini 3.7 Flash aims to become developers’ new companion
Software development is one of the main areas of improvement for Gemini 3.7 Flash. Google notably announces improvements in the model’s ability to understand codebases, identify errors, debug applications and solve problems requiring multiple reasoning steps. The goal is no longer simply to generate a few lines of code from an instruction, but to enable the model to take part in much more complete development workflows, in which it must understand an existing environment before proposing and then applying changes.¹
The results shared at launch illustrate this progress. On FrontierCode 1.1 Main, Gemini 3.7 Flash reaches 43.6%, compared with 34.4% for Gemini 3.6 Flash. The gap is even wider on DeepSWE v1.1, which focuses on software development capabilities, where the score rises from 49% to 65.3%. These are still benchmarks and will have to be tested against real-world use, but they show that Google is concentrating a significant share of its efforts on the technical skills developers are looking for.
Web development is also improving. On WebDev Arena, Gemini 3.7 Flash achieves an Elo score of 1,588, compared with 1,538 for its predecessor. The model is reportedly more accurate when it has to reproduce an interface from an existing design or build a working application while limiting the number of iterations required.²
Google no longer wants to limit Flash to code
The name “Flash” might suggest that Google prioritizes speed above all at the expense of depth. Yet this new generation is precisely trying to broaden its scope. Gemini 3.7 Flash improves on tasks that require more specialized knowledge and reasoning, particularly in finance, law, document analysis and the biosciences.
This progress shows notably on GDP.pdf, a benchmark dedicated to the analysis of complex documents, where the model reportedly reaches 34%, compared with 22% for Gemini 3.6 Flash. On AutomationBench, which focuses more on an AI’s ability to carry out automated professional tasks, Gemini 3.7 Flash climbs from 17% to 30.4%.¹ These results are especially important for Google, because they move Flash closer to a model capable of working on complete business processes rather than functioning solely as a conversational assistant.
This direction reflects a broader shift in the market. Companies are now looking less for a chatbot capable of producing an impressive answer than for a system that can understand documents, use tools, leverage data and pursue a task over several steps. Gemini 3.7 Flash is positioned squarely on this transition.
AI agents become a priority battleground
The other major ambition of Gemini 3.7 Flash therefore concerns agentic AI. An AI agent must be able to break down a goal into several operations, use different tools and progressively check the result obtained. This requires reasoning capabilities, but also speed and a cost low enough to multiply calls to the model without making each workflow economically prohibitive.
Google is thus highlighting Gemini 3.7 Flash’s ability to orchestrate complex tasks and to operate in environments where multiple agents work together. The model can, for example, help build interactive web experiences by coordinating different operations, or be integrated into pipelines combining code generation, visual creation and automation.
This shift shows how much the competition between Google, OpenAI and Anthropic is changing in nature. The question is no longer just which model provides the best answer to a given question, but which model can most efficiently carry out a complete task, with several steps, several tools and possibly several specialized agents.
Gemini Spark also benefits from this ramp-up
The progress made by Gemini 3.7 Flash should not remain confined to developer-facing interfaces. Google is also using this new generation in Gemini Spark, its agentic environment designed to automate various personal and professional tasks.
In Google Workspace, the goal is notably to enable these agents to search for information scattered across multiple files, prepare documents, draft emails or update various content while limiting manual intervention. Gemini 3.7 Flash brings a significant advantage here: a model that performs better on long, agentic tasks can, in theory, reduce the number of errors and interruptions during a workflow.
This integration also illustrates one of Google’s main competitive advantages. The company does not just own its own models; it also controls a vast software environment that includes Gmail, Drive, Docs, Sheets, Chrome, Android and Google Cloud. Every improvement to Gemini can therefore potentially be rolled out across an ecosystem used daily by hundreds of millions of people and businesses.
Aggressive pricing to accelerate adoption
Performance, however, is only part of the strategy. Google is also using price as a lever to attract developers and companies that need to run large volumes of requests. The promotional price announced for Gemini 3.7 Flash is set at $0.75 per million input tokens and $3.75 per million output tokens until the end of 2026.²
This positioning is particularly important for agentic applications. A conventional chatbot may require a single call to the model to produce an answer, whereas an agent can carry out dozens of successive interactions with an LLM to reach its goal. A difference of a few dollars per million tokens can therefore become significant when a system is deployed at scale.
For developers, Gemini 3.7 Flash is part of the Gemini ecosystem and the development environments offered by Google, notably Gemini API, Google AI Studio and Vertex AI, depending on the deployment and availability terms chosen by the company. The Gemini family has historically been offered to developers through Google AI Studio and Vertex AI, with variants specifically optimized for API use.³
Three weeks between two models: Google sets a spectacular pace
Beyond the performance of Gemini 3.7 Flash, it is perhaps the speed of its launch that is the most revealing piece of news. Gemini 3.6 Flash had barely been unveiled when its successor was already arriving. Google appears to be adopting a continuous iteration approach, in which Flash models serve as a platform for quickly incorporating technical optimizations and developer feedback without waiting for a major generation.
This strategy profoundly changes how companies need to approach the adoption of artificial intelligence models. Choosing an LLM is no longer necessarily a technology decision that holds for several years. Performance, pricing and features can change within a few weeks, forcing organizations to build architectures flexible enough to switch models or regularly compare several providers.
This cadence also raises a broader question: how far can the acceleration of development cycles go? As Google, OpenAI, Anthropic and Chinese labs shorten their timelines, the ability to seriously evaluate each new generation becomes almost as important as the ability to develop it.
The ethical stakes: more autonomous agents demand more oversight
Gemini 3.7 Flash is advancing precisely in the areas where governance issues are becoming most sensitive: programming, automation and autonomous task execution. When a model merely suggests text, an error can usually be detected before it is used. When an agent can modify a file, work on code or trigger a series of actions, the consequences of an error become potentially much more serious.
The first issue therefore concerns human oversight. The faster and more autonomous agents become, the more precisely organizations must define what they can access, modify or execute. Permissions, action traceability, approval mechanisms and the ability to interrupt a workflow must become core elements of agentic architecture, not mere options added after deployment.
Reliability also remains a central issue. The progress seen on benchmarks does not mean that Gemini 3.7 Flash is free of hallucinations, reasoning errors or incorrect code. Since the first generations of Gemini, Google has said it applies specific post-training and evaluation phases to improve the safety and alignment of its models.³ But the growing integration of these systems into professional environments requires continuous validation, particularly when they handle sensitive data or critical processes.
Finally, the frequency of new releases itself raises a governance question. A company can hardly conduct a full risk analysis if the model it has just evaluated is replaced a few weeks later. The pace of innovation will therefore gradually need to be matched by more robust standards for documenting the capabilities, limitations and changes introduced with each generation.
With Gemini 3.7 Flash, Google wants to win on three fronts
Gemini 3.7 Flash perfectly illustrates the new phase of global competition in artificial intelligence. Google is no longer simply trying to produce the model with the best absolute performance. The company wants to simultaneously improve coding, agentic capabilities and price-performance, three criteria that have become decisive for developers and organizations deploying AI at scale.
The gains announced on FrontierCode, DeepSWE, WebDev Arena and AutomationBench show a significant improvement over Gemini 3.6 Flash. But the real test will begin when developers measure this performance against their own production environments. If the benchmark gains are confirmed and the promotional pricing actually reduces the cost of agentic workflows, Gemini 3.7 Flash could quickly become one of the most attractive models in the Google ecosystem. One thing is already certain: with just three weeks between Gemini 3.6 and Gemini 3.7 Flash, Google has once again accelerated the pace of an AI race that was already moving at breakneck speed.
How does Gemini 3.7 Flash work?
Gemini 3.7 Flash is an artificial intelligence model developed by Google to combine fast execution, programming capabilities and agentic workflow automation. It succeeds Gemini 3.6 Flash just three weeks after that model’s launch and retains the hallmark positioning of the Flash line: offering a model powerful enough to handle complex tasks while keeping latency and cost suited to applications that require many calls to the model. This approach is aimed in particular at developers and companies that want to integrate artificial intelligence into large-scale production environments.
Gemini 3.7 Flash builds on the reasoning and tool-use capabilities of the Gemini family, with specific optimizations for software development. The model can analyze a codebase, understand a problem, generate a solution, identify certain errors and take part in debugging operations. The progress announced by Google shows notably on FrontierCode 1.1 Main, where Gemini 3.7 Flash reaches 43.6% versus 34.4% for Gemini 3.6 Flash, and on DeepSWE v1.1, where performance rises from 49% to 65.3%. Web development also benefits from this evolution, with an Elo score of 1,588 on WebDev Arena versus 1,538 for its predecessor.
Gemini 3.7 Flash is also designed to operate within agentic AI architectures. In this type of environment, the model no longer simply answers an isolated request. It can take part in a sequence of actions, use different tools, draw on information from multiple sources and contribute to the execution of complex professional tasks. This focus shows notably on AutomationBench, where its score reaches 30.4% versus 17% for Gemini 3.6 Flash. The model also improves in document analysis and in fields requiring specialized knowledge, with announced gains in finance, law and the biosciences.
This combination of performance and cost is central to Gemini 3.7 Flash’s positioning. Google aims to make the model affordable enough to be called multiple times within a single agentic workflow. The promotional price announced through the end of 2026 is $0.75 per million input tokens and $3.75 per million output tokens. This strategy directly targets applications in which an agent must make many calls to the model before completing a task.
- Advanced software development: generating, understanding, fixing and debugging code in complex workflows
- Agentic capabilities: taking part in action chains involving multiple steps, tools and information sources
- Professional automation: improved execution of business tasks and automated workflows
- Web development: generating interfaces and applications with greater fidelity to the instructions and designs provided
- Document analysis: processing complex documents and extracting information in professional contexts
- Specialized reasoning: announced improvements in demanding fields such as finance, law and the biosciences
- Google ecosystem: integration into Gemini environments and tools for developers and businesses
- Optimized pricing: cost positioning aimed at applications that make many calls to the model
- The announced performance is based partly on benchmarks and has yet to be confirmed in real-world professional environments
- Gemini 3.7 Flash can still produce factual errors, hallucinations or incorrect code that require human validation
- Agentic workflows can multiply calls to the model and quickly increase token consumption despite a lower unit price
- Performance can vary depending on code complexity, context length, the tools used and the domain involved
- Automating professional tasks requires rigorous management of permissions and of the data agents can access
- The proliferation of Gemini versions forces companies to regularly reassess the performance, costs and compatibility of their applications
- Sensitive or critical uses still require oversight, traceability and human validation mechanisms
Learn more
The arrival of Gemini 3.7 Flash just a few weeks after the previous generations illustrates the particularly rapid pace at which Google is evolving its models, with growing attention paid to coding and agentic workflows. On a directly related topic, read our article “Google accelerates with Gemini 3.5 Flash, an AI capable of reasoning and acting on its own”, which traces this evolution toward models that combine reasoning, autonomous execution and integration into AI agent systems.
References
1. Google. (2026). Gemini 3.7 Flash: Advancing Coding and Agentic Workflows. Google / Gemini.
2. Boursorama. (2026). Google dévoile le modèle d’IA Gemini 3.7 Flash pour le codage et les flux de travail des agents.
https://www.boursorama.com/bourse/actualites/google-devoile-le-modele-d-ia-gemini-3-7-flash-pour-le-codage-et-les-flux-de-travail-des-agents-d3495698886261c20859e4faa9363258
3. Google DeepMind. (2024). Gemini: A Family of Highly Capable Multimodal Models.
https://deepmind.google/gemini/gemini_1_report.pdf
4. MacGeneration. (2026). À peine trois semaines après la version 3.6, Google sort déjà Gemini 3.7 Flash.
https://www.macg.co/intelligence-artificielle/2026/08/peine-trois-semaines-apres-la-version-36-google-sort-deja-gemini-37-flash-310385