Model Performance

Twitter

Facebook

Copy Link

Key Facts

Key PlatformsLMSYS Arena, Open LLM Leaderboard

Key BenchmarksMMLU, HumanEval, GSM8K, ARC

Top News

Microsoft replaces OpenAI image models with its own AI technology across PowerPoint and Bing

Microsoft is swapping out OpenAI's image-generating models for its own Microsoft AI technology in PowerPoint and Bing. The shift delivers 85% cost savings in PowerPoint while offering faster performance and higher quality, according to AI model chief Mustafa Suleyman. The move signals Microsoft's push toward self-sufficiency in AI as its models now power more than half of all Microsoft products.

Technology2 sources

Microsoft replaces OpenAI image models with its own AI technology across PowerPoint and Bing

Technology2 sources

AI scores perfect 100% at International Mathematical Olympiad, matching elite human performance

Artificial intelligence has reached a milestone by scoring 100% at the International Mathematical Olympiad for the first time. Chinese tech companies Huawei and Xiaohongshu led the breakthrough, with their AI models matching the performance of just seven human contestants who achieved perfect scores out of 666 participants in Shanghai.

Technology3 sources

AI scores perfect 100% at International Mathematical Olympiad, matching elite human performance

Technology3 sources

Alibaba Qwen3.8 AI Model Claims Second Place Behind Fable 5 in Competitive Launch

Alibaba unveiled Qwen3.8-Max-Preview with 2.4 trillion parameters at the World Artificial Intelligence Conference in Shanghai, claiming it ranks second only to Anthropic's Fable 5. The launch comes days after Moonshot AI released Kimi K3 with 2.8 trillion parameters, pushing Chinese AI companies into the multi-trillion-parameter tier amid ongoing geopolitical tensions over semiconductor access.

Technology7 sources

Alibaba Qwen3.8 AI Model Claims Second Place Behind Fable 5 in Competitive Launch

Technology7 sources

Google releases faster Gemini 3.6 Flash and cybersecurity models as 3.5 Pro stays delayed

Google DeepMind unveiled three new Gemini models including the efficiency-focused Gemini 3.6 Flash, which uses 17% fewer tokens and costs less than its predecessor. The company also introduced Gemini 3.5 Flash Cyber for cybersecurity and Flash-Lite for high-volume tasks. However, the flagship Gemini 3.5 Pro remains delayed despite promises of a June release, while Google has already begun pre-training for Gemini 4.

Technology33 sources

Google releases faster Gemini 3.6 Flash and cybersecurity models as 3.5 Pro stays delayed

Technology33 sources

Pichai defends Google AI strategy as Gemini delay fuels investor concerns about AI race position

Alphabet CEO Sundar Pichai mounted a vigorous defense of Google's AI strategy during the company's earnings call, pushing back on concerns that the tech giant is losing ground in AI race. The Gemini 3.5 Pro delay, originally slated for June, has raised questions about Google's competitiveness in AI coding capabilities. Despite 82% cloud growth and strong revenue, investor concerns about AI persist as Alphabet raises its 2026 capex guidance to $205 billion.

Business14 sources

Pichai defends Google AI strategy as Gemini delay fuels investor concerns about AI race position

Business14 sources

Microsoft trains salespeople to criticize OpenAI and Anthropic as it pushes in-house AI models

Microsoft is coaching its sales teams to negatively compare AI products from OpenAI and Anthropic to its own models, marking a shift from longtime partnerships. CEO Satya Nadella criticized Anthropic Fable AI restrictions while executives positioned Microsoft's in-house models as faster, more accurate, and cost-effective alternatives despite still using partner models in Copilot.

Business7 sources

Microsoft trains salespeople to criticize OpenAI and Anthropic as it pushes in-house AI models

Business7 sources

Gemini Spark gets 50% faster with expanded Google Workspace editing and parallel processing

Google has rolled out significant upgrades to Gemini Spark, making the AI agent over 50% faster while expanding its capabilities across Google Workspace apps. The personal agent can now edit Google Docs, manage shared spreadsheets and presentations, and process multiple data sources simultaneously. Google AI Ultra subscribers worldwide gain access, while AI Pro members may soon join them.

Technology2 sources

Gemini Spark gets 50% faster with expanded Google Workspace editing and parallel processing

Technology2 sources

ZoomInfo Launches GTM Bench, First AI Benchmark Designed for Real-World Sales and Marketing Work

ZoomInfo unveiled GTM Bench, a new benchmark system that evaluates AI agents and large language models on actual go-to-market tasks. Unlike traditional AI benchmarks that test reasoning in controlled environments, GTM Bench scores systems on two critical measures: how much work they complete and whether outputs are grounded in verifiable data. In initial testing, ZoomInfo's GTM.AI achieved a GTM Bench Index of 77, significantly outperforming competitors like Apollo and Exa.

Technology2 sources

ZoomInfo Launches GTM Bench, First AI Benchmark Designed for Real-World Sales and Marketing Work

Technology2 sources

OpenAI seeks investment banking expert to train ChatGPT on Wall Street tasks

OpenAI posted a job listing for a Subject Matter Expert in Investment Banking, signaling its ambition to train ChatGPT for complex financial transactions including mergers, acquisitions, and valuations. The role, offering up to $205K plus equity, aims to define quality standards for AI-assisted banking work and improve model performance on high-stakes financial tasks.

Business2 sources

OpenAI seeks investment banking expert to train ChatGPT on Wall Street tasks

Business2 sources

Large language models can predict social science experiments but can't replace human insight

A Harvard study reveals GPT-4 can forecast outcomes of social science experiments involving nearly 120,000 participants with surprising accuracy. While large language models show promise for predicting human behavior in surveys, researchers warn that synthetic respondents aren't substitutes for real people and may create illusions of understanding without genuine insight into cognitive processes.

Science2 sources

Large language models can predict social science experiments but can't replace human insight

Science2 sources

NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Blueprint

NVIDIA Nemotron 3 Ultra delivers top accuracy among open models on LangChain's Deep Agents benchmark, achieving business task parity with leading closed models at 10x lower inference cost. The NemoClaw for LangChain Deep Agents blueprint combines an open-weight model, agent harness, and secure runtime to help enterprises build, evaluate, and deploy advanced AI agents they can fully customize and control.

Technology2 sources

NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Blueprint

Technology2 sources

Claude Fable 5 dominates Google's revamped Android AI dev benchmark with 84.5% accuracy

Google has overhauled Android Bench, its AI model benchmark for Android development, revealing Claude Fable 5 leads with 84.5% accuracy. The updated framework adds eight new models and switches to Harbor for easier testing. Despite the improvements, Google's own Gemini models lag behind competitors in Android coding tasks.

Technology2 sources

Claude Fable 5 dominates Google's revamped Android AI dev benchmark with 84.5% accuracy

Technology2 sources

Nomagic deploys AI brain in warehouse robots, halves human intervention in live operations

Warsaw-based Nomagic has deployed a vision-language-action model into live warehouse operations with paying customers, cutting robot-caused human interventions by roughly half. The company's new AI lab, led by former Google DeepMind researcher Markus Wulfmeier, is pursuing task-specific mastery before generality—a stark contrast to the industry's race toward general-purpose robot brains.

Technology2 sources

Nomagic deploys AI brain in warehouse robots, halves human intervention in live operations

Technology2 sources

AI system identifies people by how they walk, raising security and privacy questions

A new AI system called SKDMap-Net can identify individuals by analyzing their walking patterns using skeletal keypoints and joint movements. The technology achieved 95.8% accuracy on major datasets and works even when faces are obscured or cameras are far away. While it promises better long-range security checks for law enforcement, the behavioral biometrics approach raises significant privacy concerns about tracking public movement.

Science2 sources

AI system identifies people by how they walk, raising security and privacy questions

Science2 sources

AI keeps mistaking random patterns for alien life, and scientists say that's a serious problem

Researchers at Michigan State University discovered that AI systems designed to detect extraterrestrial biosignatures can be fooled 100% of the time by false positives. The study reveals a critical vulnerability in AI systems used for space exploration, raising concerns about autonomous Mars rovers and deep-space telescopes making life-detection decisions without human oversight.

Science2 sources

AI keeps mistaking random patterns for alien life, and scientists say that's a serious problem

Science2 sources

Prev12 3 4…22 Next