Google blows away competition with powerful new Gemini 3 model
Gemini 3 is the hyperscaler’s most powerful model yet and state of the art on almost every AI benchmark going
Google has officially unveiled Gemini 3 Pro, its new state of the art LLM with record-breaking scores across almost every AI benchmark.
The new model is intended to improve every Google service that uses Gemini, including its dedicated app, coding tools, and AI in search.
Google stated that Gemini 3 Pro is much better at handling requests in their intended context and providing useful answers that don’t resort to flattery.
The model debuted in the number one spot across text, WebDev, and vision in LMArena, with Google having claimed Gemini 3 Pro is the “best model in the world for complex multimodal understanding”.
Across multimodal reasoning benchmarks, Gemini 3 Pro was found to consistently outperform competition such as GPT-5.1 and Claude Sonnet 4.5.
In MMMU-Pro, for example, the model scored 81% versus GPT-5.1’s 76% and Claude Sonnet 4.5’s 68%.
ARC-AGI-2 is a rigorous benchmark for testing the capability of AI model reasoning across a series of abstract visual puzzles. Easy for humans to complete but difficult for today’s LLMs, it’s considered a true challenge for frontier models.
Sign up today and you will receive a free copy of our Future Focus 2026 report - the leading resource for IT decision-maker insight on priorities and investment areas in AI, security and more.
Gemini 3 Pro scored 31.1% in tests, far in excess of GPT 5.1’s 17.6% and Claude Sonnet 4.5’s 13.6%.
Gemini 3 is a game changer for Google shops
In other areas, the performance gap between Google’s model and the competition is even more stark.
Gemini 3 Pro scored a new record score of 23.4% at MathArena Apex, a benchmark that tests LLMs on their ability to solve mathematical problems upon which they weren’t trained, compared to just 1% by GPT-5.1 and 1.6% by Claude Sonnet 4.5.
Although the model doesn’t beat the latest version of Claude at the most common coding benchmarks, it greatly improves upon Gemini 2.5 Pro.
Google has also emphasized how its reasoning and visual understanding helps Gemini 3 Pro to do more with the coding capabilities it has, for quicker overall resolution of common developer tasks.
In a demo, Google showed how Gemini 3 Pro could turn an image of a chessboard into an interactive game, as well as a back-of-the-napkin sketch of a website into a functioning page.
Gemini 3 Deep Think, the multi-step reasoning variant of the new model, improves upon its benchmarks at the cost of far slower responses.
In Humanity’s Last Exam, an intense benchmark designed to press LLMs with 2,500 difficult questions across a wide range of subject areas, Gemini 3 Deep Think scores 41%, compared to Gemini 3 Pro’s 37.5% and GPT-5 Pro’s 30.7%.
Gemini 3 Deep Think is still undergoing safety tests but will become available to Google AI Ultra subscribers within the next few weeks.
Gemini 3 Pro was trained using Google’s tensor processing units (TPUs), dedicated hardware for AI training and inference that Google cites as key to its efforts to develop AI sustainably.
Google Antigravity
Alongside the much-anticipated launch of Gemini 3 today, Google also revealed Google Antigravity.
The product is described as a new ‘agent-first’ IDE, with a focus on elevating developers into a manager of AI agents.
Using Gemini 3’s agentic coding features, as well as its reasoning and tool use, Antigravity is intended to allow developers to set agents multi-step, complex coding tasks and receive evidence proving the job has been completed to a high standard.
In a demo, Google showed how a developer could use the tool to build a flight lookup web app, returning flight data based on a flight number provided by the user.
The tool is then capable of creating an implementation plan, writing the code, and then testing the app by opening the Chrome browser. Finally, it provides the user with screenshots of its tests, which the user can critique in order to immediately update the UI of the finalized app.
These features are powered by a combination of Gemini 3, Gemini 2.5 Image (better known as Nano Banana), and Google 2.5 Computer Use.
In another demo, Google showed how enterprise developers can delegate to multiple ‘background agents’ within Antigravity, without needing to stay in the same chat window while tasks are being completed.
Google added that Antigravity is capable of supporting Claude 4.5 Sonnet in addition to its own models.
Antigravity is now in public preview for free, with support for other models including Claude Sonnet 4.5 and OpenAI’s GPT-OSS.
Enterprises with the Vertex AI and Gemini Enterprise subscriptions have access to Gemini 3 from today and the model is also available in the Gemini app, the Gemini API within AI Studio, Antigravity, and Gemini CLI.
The model will be priced at $2 per million input tokens and $12 per million output tokens, for prompts under 200,000 tokens.
Prompts over this limit will cost $4 per million input tokens and $18 per million output tokens.
Make sure to follow ITPro on Google News to keep tabs on all our latest news, analysis, and reviews.
MORE FROM ITPRO
Rory Bathgate was Features and Multimedia Editor at ITPro until 2026, overseeing all in-depth content and case studies. He also co-hosted the ITPro Podcast with Jane McCallion, swapping a keyboard for a microphone to discuss the latest learnings with thought leaders from across the tech sector.
In his free time, Rory enjoys photography, video editing, and good science fiction. After graduating from the University of Kent with a BA in English and American Literature, Rory undertook an MA in Eighteenth-Century Studies at King’s College London. He joined ITPro in 2022 as a graduate, following four years in student journalism. You can contact Rory on LinkedIn.
-
NetApp ticks another hyperscaler box with Oracle Cloud Infrastructure dealNews The deal with Oracle brings the number of native hyperscaler integrations to four at NetApp
-
Integrating probabilistic programming in enterprise risk managementProbabilistic risk modeling can transform statistical predictions into a programming language, making outputs more dynamic
-
Huawei is positioning itself as a viable alternative to big US tech companies. Can it pull it off?Opinion With Nvidia and AMD excluded from the Chinese AI market, Huawei is building a powerful reputation at home and abroad
-
Google targets AI cost efficiency with new FinOps features for Gemini EnterpriseNews New cost control features and subscription options for Gemini Enterprise look to bring down spiralling AI costs
-
“Enterprises are becoming much more rigorous about the economics of AI”: Snowflake wants to help you cut AI costs by choosing the right model for the right taskNews New dynamic model routing capabilities aim to help customers box clever when it comes to their AI model choice
-
Google just spent $10 million in an auction for Spirit Airlines data – 100 million emails, 500 million Microsoft Teams chats, and 30 million lines of code will be used to improve AI models and productsNews AI developers are struggling to find enough data to train their models, sparking a bidding war for failed Spirit Airlines deidentified data
-
Microsoft has joined the growing list of companies cracking down on ‘tokenmaxxing’News The company is updating internal guidance to reduce rising costs
-
Oracle integrates Google's Gemini AI models into enterprise appsThe deal aims to give broader access to Gemini models that can support AI agents and accelerate development
-
‘AI cost management has the same problems that cloud had’: Enterprises are still facing huge AI bills thanks to ‘tokenmaxxing’ – that means FinOps practices are more important than everNews With firms facing surging AI bills, FinOps techniques are more important than ever
-
‘We are now seeing MAI models outperform general-purpose frontier models’: Microsoft CEO Satya Nadella touts in-house models to cut spiralling AI costs – and reduce growing reliance on frontier labsNews The Microsoft chief says pricey frontier models don't have to be used for every task, and its own in-house MAI models could be the key to reducing costs.