Google's new Gemini 3.6 Flash shows strong results in document drafting and review while completing tasks an average of 12% faster, according to Harvey’s Niko Grupen.
Although Google entered this year's model race after Anthropic and OpenAI, the company is now pushing harder to compete with a new Gemini lineup focused on speed, efficiency and affordability. The models are designed to challenge leading rivals, including Chinese AI companies, across both pricing and performance.
Google says the new release is built to give developers greater token efficiency, reduced latency and more dependable task execution, particularly for high-volume AI agents operating in production environments.
The launch includes Gemini 3.5 Flash Cyber, introduced around a month after Anthropic's Mythos and Fable security models experienced an unexpected freeze followed by a relaxation of restrictions.
Google is positioning three separate cost-efficient models to compete in this space, with each one tailored to different requirements around throughput, security, and operating costs.
The new competitors
Building on the Gemini 3.5 Flash series, Google is expanding its Flash lineup with models designed to balance performance, speed and cost efficiency.
- Gemini 3.6 Flash: Brings stronger coding, knowledge-work and multimodal capabilities while using 17% fewer output tokens on the Artificial Analysis Index. It also comes with a lower cost per output token.
- Gemini 3.5 Flash-Lite: The fastest and most affordable model in the 3.5 class, Flash-Lite reaches 350 output tokens per second on the Artificial Analysis Index and is designed to improve agentic workflows compared with earlier generations.
- Gemini 3.5 Flash Cyber in CodeMender: This release combines a specialized cybersecurity model with Google’s CodeMender security agent, targeting competitive performance at the frontier of AI-powered cyber defense.
Google is also testing Gemini 3.5 Pro with selected partners before making the model more widely available. At the same time, the company has begun pre-training Gemini 4, signaling its continued push toward the next generation of AI models.
Together, these releases represent Alphabet's latest effort to compete with Anthropic in automated code security while expanding the use of AI agents across high-volume production environments.
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are currently available through Google AI Studio, Android Studio, Google Antigravity, Gemini Enterprise and the Gemini App.
Enhanced Coding and Computer Capabilities
According to Artificial Analysis data, Gemini Flash is positioned as a lower-cost alternative to comparable models from Anthropic, OpenAI and Chinese AI companies. The model costs less per task than GPT-5.6 Terra Max, Kimi K3 and Qwen 3.7 Max, all of which target high-volume production use cases.
Gemini 3.6 Flash also shows stronger accuracy while making fewer unnecessary code changes. On DeepSWE, it achieved 49%, compared with 37% for the previous result. Its performance in machine learning research improved as well, reaching 63.9% on MLE Bench versus 49.7%.
The model has also made progress in computer-use tasks, scoring 83% on OSWorld-Verified compared with 78.4%. The capability is available as a built-in client-side tool through the Gemini API and Gemini Enterprise.
In knowledge-based tasks, Gemini 3.6 Flash surpassed Gemini 3.5 Flash on GDPval-AA v2, scoring 1,421 against 1,349. Companies including Hebbia and Harvey have reported that the model performs well in areas such as document analysis, chart interpretation, and report creation.
Niko Grupen, Head of Applied Research at Harvey, said Gemini 3.6 Flash performs particularly well for document drafting and review in areas such as capital markets and corporate M&A.
He added that the newer model delivered substantial benchmark improvements over its predecessor while completing tasks an average of 12% faster, highlighting its gains in efficiency as well as performance.
Scaling Agentic Workloads
Gemini 3.5 Flash-Lite targets low-latency, high-through workloads such as agentic search and document processing. With an output speed of 350 tokens per second, Google positions it as the fastest model in the 3.5 series.
At $0.30 per 1 million input tokens and $2.50 per 1 million output tokens; it is designed to deliver a competitive balance between cost and performance. It also surpasses 3.1 Flash-Lite across different thinking levels, supporting more efficient scaling.
Developers can use minimal thinking for large-scale, high-volume tasks or increase the reasoning level when handling complex, multi-step subagent workflows. Computer use is integrated as a built-in tool, allowing agents to perform tasks across different surfaces.
The model records significant gains in coding and agentic workloads, scoring 54% versus 31% on Terminal-Bench 2.1. It also performs better on long-context tasks on GDM-MRCR v2, with 72.2% versus 60.1%, and real-world execution on GDPval-AA v2, where it scores 1,140 compared with 642.
Flash-Lite also outperforms 3 Flash on SWE-Bench Pro, scoring 54.2% versus 49.6%, and on OSWorld-Verified, reaching 74% compared with 65.1%.
Beyond benchmarks, the model can extract product attributes from large ecommerce datasets and synthesize information. Working alongside Gemini 3.6 Flash as the master agent, it can rapidly produce 25 different web design concepts.
The model can also handle receipt translation at scale using multimodal understanding and create games by testing and refining multiple approaches. Early customers have praised its speed, intelligence, and cost efficiency.
Advancing AI Code Defense
Google is increasingly focused on the reality of modern cyber threats, where AI models can identify security vulnerabilities faster than organizations can address them.
To respond to this challenge, the company has developed Gemini 3.5 Flash Cyber, a specialized version of 3.5 Flash fine-tuned to detect and remediate cybersecurity vulnerabilities while keeping token costs lower.
Through CodeMender, multiple Gemini 3.5 Flash Cyber agents can work together and combine their findings into a single report. The approach delivers competitive frontier-level results on the CyberGym benchmark and helps Google close the cybersecurity performance gap with Anthropic, which gained an early advantage with its Mythos model.
Google plans to make the model available exclusively to governments and trusted partners through a limited CodeMender pilot. The program is intended to give cybersecurity defenders more time to identify and patch critical vulnerabilities before attackers can exploit them.
Meanwhile, Chinese AI companies continue to strengthen their position. Moonshot AI has restricted Kimi K3 subscriptions because of capacity limitations, while Alibaba has hinted at Qwen 3.8 Max. Competing at this level also requires enough computing infrastructure to operate AI models at scale.
Google has an advantage in this area through its custom chips, cloud infrastructure and approach to co-designing AI models with the hardware that runs them.
By combining speed, affordability and performance, the new Flash lineup strengthens Google's push to compete for large-scale production of AI agent workloads as developers look for faster and more cost-efficient alternatives to established rivals.