Google Launches Gemini 2.5 Pro with "Deep Think" Reasoning Mode and Gemini 3.5 Flash "Computer Use" API

1. Introduction: Gemini 2.5 Pro with "Deep Think" Reasoning

Google has officially announced the launch of **Gemini 2.5 Pro**, introducing a highly anticipated **"Deep Think"** reasoning mode. Similar to OpenAI's reasoning architecture and Anthropic's multi-step agents, Deep Think allows the model to dedicate extra test-time compute to solve complex, multi-layered problems in fields like computer science, mathematical proofs, and scientific analysis before generating its output.

This release comes as a direct response to the intense competition in the frontier AI model space. With startups and cloud hyperscalers investing trillions of dollars in physical infrastructure—as analyzed in our article on South Korea's landmark $1 trillion AI and semiconductor strategy—model capabilities are shifting rapidly toward specialized reasoning and autonomous task execution.

2. The 2-Million Token Window and Benchmark Achievements

Gemini 2.5 Pro retains its industry-leading **2-million token context window**, allowing users to upload entire codebases, hours of audio-visual content, or hundreds of research papers. However, with the integration of "Deep Think," Google reports significant improvements in benchmark tests:

• GPQA (Graduate-Level Google-Proof Q&A): Scores rose by 14% compared to Gemini 2.0 Pro.
• SWE-bench Verified: Resolving real-world software engineering issues saw a 19% performance boost.
• Math Olympiad Benchmark (AIME): Shows a distinct leap in symbolic reasoning accuracy.

3. Gemini 3.5 Flash: Public Preview of "Computer Use" API

In addition to Gemini 2.5 Pro, Google rolled out a public preview of the **"Computer Use" API** for its lightweight **Gemini 3.5 Flash** model. This tool allows the AI model to perceive and interact directly with computer interfaces—simulating mouse movements, clicks, keystrokes, and scroll actions.

This release aims to power autonomous software agents that can automate tasks like filling out CRM reports, testing web applications, or organizing files across spreadsheets. While Anthropic has led with its own Computer Use APIs, Google's integration offers extremely low latency due to Gemini 3.5 Flash's optimized inference pipelines. For reference on chip-level cost optimizations, read about OpenAI's custom "Jalapeño" inference chip.

4. Google's Nano Banana Model & Personalized Images

For consumer-facing features, Google has introduced free personalized image generation within the Gemini app. Powering this feature is the new **"Nano Banana"** image generation model. Nano Banana runs locally on compatible mobile devices and works in tandem with Google Photos to generate stylized, personalized visual assets for users at zero API cost.

5. The Infrastructure Constraints Delaying Gemini 3.5 Pro

Despite these releases, industry analysts have noted the absence of **Gemini 3.5 Pro**, which Google initially planned to launch by mid-2026. Sources indicate that the launch has been delayed indefinitely due to severe computing infrastructure bottlenecks and GPU constraints. Google is reportedly prioritizing internal compute resources for central search features and limiting compute access to third-party clients. This infrastructure crisis mirrors our recent reporting on compute limits forcing Meta to shift onto internal models, as well as the BIS warning regarding the $1 trillion AI infrastructure bubble.

6. Comparison: Gemini 2.5 Pro vs. Competitors

7. Frequently Asked Questions (FAQ)