A short daily roundup of relevant AI news from roughly the last 24–48 hours. We already covered the weekend pacing and guardrails wave (Amodei’s essay, Altman’s clarification, Trump on Truth Social) in the September 14 roundup — we are not repeating those; below is the continuation plus new product announcements.
Google: Gemini 3.8 Live and 3.8 Live Extended Thinking are here. On September 15 Google officially introduced two new near-real-time speech/dialogue models. 3.8 Live is built for scalable, cost-efficient live conversation and visual context; 3.8 Live Extended Thinking targets harder multi-step tasks, with background tool calls and live “thinking” narration. Google says Extended Thinking ranks #1 on the Artificial Analysis Speech-to-Speech Quality Index (82.6), handles 97 languages, and watermarks every AI-generated audio clip with SynthID. For developers: Gemini API and AI Studio; on the consumer side, a gradual rollout in Search Live / Gemini Live / Workspace (Docs, Gmail, Keep). Pricing per the developer post: about $0.005/min audio input and $0.018/min output. Source: Google Blog, developer audio announcement.
Clarification / continuation of our earlier post: Latest AI news — September 14, 2026 — On Tuesday OpenAI publicly confirmed it has been in talks for weeks with Anthropic and Google DeepMind on AI safety cooperation, along the lines of Hassabis’s July “Standards Body” proposal. Chris Lehane (OpenAI global affairs) spoke about this in Washington; according to TechCrunch/Bloomberg the companies are not seeking a separate antitrust exemption for safety coordination. In the same remarks Lehane said OpenAI supports the part of the bipartisan FRONTIER Act that would require independent verification organizations (IVOs) for large frontier labs. Source: CNBC, TechCrunch, Washington Examiner.
Musk: rival labs should test each other’s models. At the All-In Summit (Los Angeles) Elon Musk proposed that leading AI companies — SpaceXAI/xAI, OpenAI, Anthropic, Google, Meta, and “three or four leading Chinese companies” — run one another’s safety “test harnesses” on their models before public release, rather than grading their own homework. He acknowledged rivals have not agreed yet; he thinks China may also find this more acceptable than a slowdown. This is a new proposal relative to his weekend support for pacing — we already covered the Grok 4.8 RL timeline yesterday. Source: CNBC, summary: TeslaNorth.
TotalEnergies + Mistral: a 3-year, more than €100 million frontier-AI program. Paris, September 15: TotalEnergies and Mistral are launching a joint program for frontier AI models to support reservoir research and reservoir engineering — more than €100 million in investment, a joint scientific lab, agentic AI, and TotalEnergies’ nearly century-long geoscience data. Patrick Pouyanné and Arthur Mensch both emphasized a European digital ecosystem and own IP. This is not a rerun of the September 8 Series D funding, but a separate industrial partnership. Source: TotalEnergies press release.