Google’s Gemini 2.5: Thinks deeper, speaks smarter, codes faster

At this year’s Google I/O, the tech giant unveiled sweeping enhancements to its Gemini 2.5 models, signaling a major leap toward its ambitious vision of building a “universal AI assistant” — one that is personal, proactive, and powerful.

Demis Hassabis, CEO of Google DeepMind, laid out the company’s ultimate goal during a press briefing: to create an AI that doesn’t just respond, but understands context, reasons across complex inputs, plans, and takes meaningful action.

Gemini 2.5 Pro Gets Smarter with 'Deep Think'

Among the biggest announcements is the experimental Deep Think mode in Gemini 2.5 Pro — Google’s most intelligent AI model to date. Deep Think enables the model to consider multiple hypotheses before forming a response, leveraging learnings from DeepMind’s work on AlphaGo and other advanced reasoning systems.

Deep Think’s performance is already drawing attention. It scored highly on rigorous benchmarks like the 2025 USA Mathematical Olympiad (USAMO) and LiveCodeBench, which tests competition-level coding. It also achieved an 84% score on MMMU, a benchmark for multimodal understanding and reasoning. These scores suggest that Gemini 2.5 Pro is not just faster — it's significantly more capable at tasks requiring deep reasoning.

The model also leads the WebDev Arena coding leaderboard with an ELO score of 1420, marking it as a top-tier performer. It dominates the LMArena leaderboard too, which ranks models based on human preference across a range of tasks.

Despite these milestones, Google is holding back broad release of Deep Think for now, opting to conduct additional safety testing with trusted testers via API access.

Gemini 2.5 Flash: Lighter, Faster, Still Powerful

While Pro takes center stage for depth and intelligence, Gemini 2.5 Flash is designed for speed and efficiency. Now significantly improved across reasoning, code, and long-context understanding, it’s “second only” to Pro in overall performance, according to Hassabis.

Flash is also more resource-efficient, using 20–30% fewer tokens — a big plus for developers aiming to reduce costs without sacrificing capability. It is available now in preview via Google AI Studio, Vertex AI, and the Gemini app, with general availability set for early June.

New Features: Audio, Emotion, and Better Thought Transparency

Google also introduced a host of new features that stretch the boundaries of what conversational AI can do:

• Native audio output and text-to-speech with multi-speaker support: Gemini can now speak in 24+ languages and adapt tone, accent, and style (even whisper).
• Affective dialogue: The models can detect emotion in a user’s voice and respond appropriately.
• Proactive audio: This allows the model to tune out background noise or distinguish between different voices in a room.
• Live API “thinking”: Supports more complex, ongoing tasks with real-time responsiveness.

Another major addition is “thought summaries”, which organize the model’s internal reasoning into structured, user-friendly formats — complete with headers, key details, and descriptions of tool usage. This not only enhances transparency but also helps developers debug and understand model behavior more easily.

Also introduced were “thinking budgets”, letting developers control the number of tokens the model can use to "think" before responding — or turn the thinking process off entirely for simpler tasks.

Deep Integration and Developer Tools

For developers, Google is rolling out native SDK support for Model Context Protocol (MCP) in the Gemini API, streamlining integration with open-source tools and custom applications.

These updates not only reflect Google’s technical progress but reinforce its strategic aim: to build an AI assistant that operates with fluidity across speech, code, planning, and reasoning.

The Road Ahead

“This is a remarkable moment in history where AI is making possible an amazing new future,” said Hassabis. With relentless iteration and careful focus on safety and capability, Google is rapidly shaping that future — one where AI isn't just helpful, but truly intelligent.