Google Launches Gemini 3 Pro: 37.5% HLE Score and Instant Search Integration

Image Credit: Jacky Lee

Alphabet’s Google released Gemini 3 on Tuesday, unveiling its most advanced frontier AI model to date and highlighting the increasingly rapid pace of foundation-model development across the industry. The launch follows Gemini 2.5 Pro, introduced earlier in 2025, marking Google’s second major upgrade of the year and underscoring the accelerating iteration cycles among leading AI developers.

Developed by Google DeepMind, Gemini 3 Pro became available immediately across Google’s primary consumer and enterprise AI interfaces, including the Gemini app, Google Search’s AI Mode, and Google’s developer tools. This same-day deployment into monetised products marks a significant change from earlier Gemini rollouts, where capabilities were typically phased in over weeks.

Key Capabilities and Technical Advances

The first model in the Gemini 3 family, Gemini 3 Pro, introduces major improvements in reasoning, multimodal understanding (spanning text, images, audio and video), and agentic capabilities for autonomous multi-step task execution. Google describes it as its most intelligent model to date, designed to interpret user intent more precisely and deliver more contextually relevant responses.

A standout element of the launch is Google’s introduction of “generative interfaces” — dynamic, interactive response layouts that present magazine-style explanations, visual diagrams, sliders, structured modules and custom UI components. These interfaces aim to make complex topics, such as scientific processes, cultural analysis or art history, more intuitive and visually rich.

A higher-reasoning variant, Gemini 3 Deep Think, remains in final safety testing. Google confirmed it will be released in the coming weeks to Google AI Ultra subscribers, following further evaluation to ensure reliability on long-form reasoning tasks.

Google also unveiled Antigravity, a new agentic coding environment powered by Gemini 3. The platform features a multi-pane workflow combining code editors, terminals, browser views and tool integrations, allowing AI agents to plan, navigate and execute development tasks across tools in real time.

Benchmark Performance

Early benchmark results place Gemini 3 Pro competitively among frontier models:

• Humanity’s Last Exam: 37.5% without external tools — one of the strongest scores reported at launch across tested models.

• SWE-bench Verified: 76.2%, placing Gemini 3 Pro alongside the top tier of models for software-engineering and code-reasoning tasks.

Background and Development Context

Gemini 3 arrives amid compressed release cycles across the AI sector. The update follows Gemini 2.5 Pro earlier this year, while major competitors have also accelerated their cadence: OpenAI released GPT-5.1 just days prior, and Anthropic launched Claude Sonnet 4.5 in late September.

Google’s faster rhythm reflects its deep investment in compute infrastructure and the breadth of its product ecosystem. With the Gemini app recently surpassing 650 million monthly active users, Google can deploy new model capabilities at global scale from day one — an advantage few competitors share.

The launch stands in sharp contrast to the original Gemini 1.0 rollout in late 2023, which was criticised for factual inaccuracies and problematic image outputs. In response, Google emphasises strengthened safety processes in Gemini 3, including deeper behavioural testing, reduced tendencies toward sycophancy, and improved resilience against adversarial and prompt-injection attempts.

Immediate Availability and Access

Gemini App

Gemini 3 Pro is rolling out globally starting today. Users can access it by selecting the “Thinking” option in the app’s model selector. Premium subscription tiers — Google AI Plus, Pro and Ultra — continue to offer expanded capabilities and higher usage limits.

Google Search (AI Mode)

In the United States, Google AI Pro and Ultra subscribers can now access Gemini 3 Pro directly through AI Mode by selecting “Thinking”. Google stated that automatic routing of more complex queries to Gemini 3 will expand in the coming weeks.

Developers

Developers can begin using Gemini 3 Pro immediately through:

• Google AI Studio

• Vertex AI

• Gemini CLI

• Google Antigravity

The Deep Think variant will be released separately following additional safety evaluations.

Industry Implications and Future Trends

Analysts point to Google’s decision to embed a frontier-scale model into Search and its wider ecosystem on launch day as a strategic demonstration of the advantages of vertical integration. Unlike competitors reliant on standalone applications, Google can activate new capabilities across billions of user touchpoints simultaneously.

The launch also reflects a broader shift toward agentic AI systems — models that not only generate content but autonomously navigate tools, plan workflows and produce adaptive, interactive outputs tailored to user needs.

Google confirmed that additional models in the Gemini 3 family will follow, continuing a trend of sub-annual model releases driven by training advances and competitive pressure. As development timelines shorten, industry observers are paying closer attention to the sustainability of rising compute demands and the adequacy of safety oversight amid accelerating progress.

License This Article

Source: Google Blog, The Verge, Story Board 18, LLM Stats, OpenAI Platform, Claude Sonnet 4.5, Techzine, Technology Magazine, eWeek, Wired

$5.00

$10.00

$20.00

$30.00

Custom Amount

Please enter an amount

$

One-Time Donation Monthly Donation Quarterly Donation Annual Donation

3% Cover the Fee

Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform! Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform!