Chinese startup DeepSeek has launched 685B-parameter models V3.2 and V3.2-Speciale, matching or surpassing GPT-5 and Gemini-3.0-Pro on major math and coding benchmarks. A 2025 technical report on Hugging Face details the DeepSeek Sparse Attention architecture, which cuts 128K-token inference costs by about 70% compared with the prior V3.1-Terminus model.
Anthropic released Claude Opus 4.5, its newest flagship model, with advances in coding benchmarks, enterprise automation and integrations across Chrome, Excel and desktop, plus new pricing, memory and safety improvements, and developer tools for long-running agentic workflows.
Google released Nano Banana Pro, the Gemini 3 Pro Image model from Google DeepMind, delivering high-fidelity visuals, multilingual on-image text, and SynthID-backed AI image verification across the Gemini app, Search, Workspace, Ads, and developer tools.
Meta introduces SAM 3 and SAM 3D, combining open-vocabulary segmentation, a large-scale hybrid annotation pipeline, creator-focused tools like the Segment Anything Playground, and new 3D benchmarks such as SA-3DAO for grounded reconstruction.
xAI introduced Grok 4.1 with major gains in reasoning, safety, emotional intelligence, and multimodal capability, plus competitive pricing and rapid API availability for enterprise use.
Google unveiled Gemini 3 and Gemini 3 Pro, its top-performing AI models, integrating them into Search and Workspace, claiming 72% answer accuracy, record reasoning and coding benchmarks, and launching the Antigravity agentic coding platform amid fierce competition and rising infrastructure costs.
OpenAI introduced GPT-5.1 Instant and Thinking with adaptive reasoning, warmer tone defaults, improved instruction following, and expanded tone presets and sliders for more natural, precise responses.
Google Maps gains Gemini AI for multi-step voice commands, U.S.-first proactive traffic alerts, landmark-based guidance, and Lens-powered local discovery on Android and iOS.
OpenAI’s Aardvark is a GPT-5–powered agentic security researcher now in private beta, benchmarked at ~92% recall on seeded “golden” repos and credited with 10 CVE-tracked disclosures.
