Google’s latest AI models compared with features and performance

Gemini 3 is designed to offer higher solutions to extra advanced questions in contrast with prior fashions. Additionally it is stated to be the very best mannequin that Google has constructed for ‘vibe-coding’, the controversial apply the place customers largely depend on AI instruments to generate code and construct software program.

In keeping with Google, its advances with Gemini 3 are mirrored within the mannequin’s efficiency throughout a number of benchmark exams. The corporate claimed that Gemini 3 outperforms its predecessor on each AI benchmark, topping the LM Enviornment leaderboard in addition to incomes high marks on Humanity’s Final Examination and GPQA Diamond.

Nevertheless, public benchmarks have been criticised as unreliable indicators of real-world AI efficiency as a result of they are often simple to sport. For example, famed AI researcher Andrej Karpathy identified that Gemini 3 refused to consider that it was 2025 since its pre-training information solely included data up until 2024. However he additionally acknowledged that his early impression of Gemini 3 was constructive.

I performed with Gemini 3 yesterday by way of early entry. Few ideas –

First I often urge warning with public benchmarks as a result of imo they are often fairly doable to sport. It comes right down to self-discipline and self-restraint of the staff (who’s in the meantime strongly incentivized in any other case) to…

— Andrej Karpathy (@karpathy) November 18, 2025

As suggestions continues to roll in over the following few weeks, let’s take a better take a look at the Gemini 3 household of fashions and what every of them has to supply.

Gemini 3

Gemini 3 is alleged to own multimodal reasoning capabilities, which means that it combines reasoning skills with imaginative and prescient and spatial understanding in addition to multilingual expertise and a one million-token context window, permitting customers to ask advanced and nuanced questions, together with prolonged ones.

For builders, Gemini 3 is able to dealing with advanced prompts and directions to render richer, extra interactive net UI. In keeping with Google, Gemini 3 is ‘distinctive’ at zero-shot era which signifies that it may well generate software program parts with out being explicitly skilled on such parts.

When it comes to use circumstances, Google stated that customers may, as an illustration, ask Gemini 3 to decipher and translate handwritten recipes in numerous languages right into a shareable household cookbook. “It will probably even analyse movies of your pickleball match, establish areas the place you possibly can enhance and generate a coaching plan for total kind enhancements,” the corporate stated.

Gemini 3 has been subjected to a number of security exams to be able to scale back sycophancy and enhance resistance to malicious immediate injection assaults, as per Google.

On the benchmark entrance, Gemini 3 topped the WebDev Enviornment leaderboard by scoring a formidable 1487 Elo. It additionally scores 54.2 per cent on Terminal-Bench 2.0, which exams a mannequin’s software use capacity to function a pc by way of terminal. It outperformed Gemini 2.5 Professional on SWE-bench Verified (76.2 per cent), a benchmark that measures coding brokers.
The mannequin additional topped the Merchandising-Bench 2 leaderboard, which exams longer horizon planning by managing a simulated merchandising machine enterprise.

Gemini 3 Professional

“Gemini 3 Professional demonstrates higher long-horizon planning to generate considerably increased returns in comparison with different frontier fashions,” Google stated.

Its responses are good, concise, and direct. “It acts as a real thought associate that offers you new methods to know data and specific your self, from translating dense scientific ideas by producing code for high-fidelity visualizations to artistic brainstorming,” the corporate added.

Geminin 3 Professional outperforms 2.5 Professional on each main AI benchmark. It topped the LMArena Leaderboard with a breakthrough rating of 1501 Elo. It obtained high scores on Humanity’s Final Examination (37.5 per cent with out the utilization of any instruments) and GPQA Diamond (91.9 per cent).

Gemini 3 Professional is alleged to be extremely succesful at fixing advanced issues throughout an enormous array of subjects like science and arithmetic. It set a new excessive rating (23.4 per cent) on MathArena Apex, a benchmark for evaluating frontier fashions on arithmetic. Its multimodal reasoning extends past textual content because the mannequin scored 81 per cent on MMMU-Professional and 87.6 per cent on Video-MMMU.

Gemini 3 Professional’s responses are additionally extra prone to be factually correct because it scored a 72.1 per cent on SimpleQA Verified.

Gemini 3 Deep Suppose

Gemini 3 Deep Suppose is an enhanced reasoning mode that pushes Gemini 3’s multimodal reasoning capabilities even additional to assist customers resolve extra advanced issues.

In testing, Gemini 3 Deep Suppose outperformed Gemini 3 Professional’s efficiency on Humanity’s Final Examination (41.0 per cent with out using instruments) and GPQA Diamond (93.8%). It additionally achieved 45.1 per cent on ARC-AGI-2 (with code execution, ARC Prize Verified), demonstrating its capacity to unravel novel challenges.

Nevertheless, Google stated that Gemini 3 Deep Suppose Mode continues to be present process security evaluations and will probably be made accessible to Google AI Extremely subscribers after gathering inputs from security testers within the coming weeks. The corporate has additionally stated it plans to launch extra fashions to the Gemini 3 collection quickly.