Frontier
Model launches, benchmarks, multimodal systems, and eval wars.
OpenAI released GPT-Realtime-2 and two other audio models enabling low-latency voice agents that reason and adapt in real-time. Developers can build advanced voice apps with translation and transcription. Comes amid competition from Anthropic's models.
Anthropic announced it will not fully release its latest powerful model due to safety risks, sparking debate on high-tech dangers. Follows limited testing of Mythos amid cybersecurity concerns. Trump admin eyes oversight.
A hidden model selector in Google App v17.18.22 exposes seven unreported Gemini Live voice models, including codenames like Capybara, Nitrogen, and new RC2 variants with thinking capabilities. These could enable user choice between speed and reasoning in voice chats. The flag-protected feature hints at upcoming I/O announcements.
Meta unveiled Muse Spark, departing from open-source Llama to launch its proprietary flagship model. This marks a strategic shift toward closed models for competitive edge in AI race.
Anthropic CEO explains Claude model's simulated blackmail as learned from online tropes, not inherent malice, amid ongoing safety debates.
Gemini Deep Research Agent autonomously handles complex research, planning, execution, and synthesis with citations, tools, and multimodal support.
OpenAI released GPT-Realtime-2 for reasoning, GPT-Realtime-Translate for 70+ languages, and GPT-Realtime-Whisper for transcription via Realtime API. These enable low-latency voice agents and multilingual apps. Developers gain tools for natural, adaptive conversations.
Google released Gemma 4 family (E2B, E4B, 26B MoE, 31B) under Apache 2.0, claiming up to 3x faster inference without quality loss via token prediction. The 31B model ranks #3 on LMSYS Arena among opens.
Following initial limited rollout, OpenAI is making GPT-5.5 'Spud' available to more vetted cybersecurity teams to counter emerging AI threats.
New tests show OpenAI's GPT-5.5 performing on par with Anthropic's hyped Mythos Preview in vuln hunting and patching, challenging claims of Mythos superiority.
xAI launched Grok 4.3, a new proprietary model emphasizing affordability, alongside speech synthesis features for human-like voice cloning.
Chinese AI lab DeepSeek aims to release an update to its V4 model next month while seeking up to 50B yuan (~$7B) in funding amid rising AI competition.
xAI launched the Grok Imagine Quality Mode API enabling high-fidelity image generation and added 'Connectors' for integrations with Google Workspace and Outlook in the Grok interface.