This brief covers AI news from 2026-08-25 UTC.
Today’s AI news: Nvidia’s investment talks, OpenAI’s price cuts, and new model releases.
Nvidia in Talks to Invest in Perplexity at $30B+ Valuation
What happened: Nvidia is discussing an investment in Perplexity as part of an equity round valuing the AI search startup at more than $30 billion, per The Information. Perplexity’s annualized revenue has risen above $750 million, driven partly by Perplexity Computer.
Why it matters: A Nvidia investment would deepen ties between the chip giant and a fast-growing agent startup, potentially shaping AI infrastructure and agent adoption.
OpenAI Cuts GPT-5.6 Sol API Prices by Over 20%
What happened: OpenAI cut developer pricing for its frontier GPT-5.6 Sol model by more than 20% for three months, effective on API and eligible plans for ChatGPT Work and Codex. New prices: $4 per 1M input tokens and $20 per 1M output tokens, down from $5 and $30.
Why it matters: Cheaper frontier tokens lower the cost of building agentic applications, intensifying price competition with Anthropic and Chinese models.
Meta Launches Muse Code Terminal Agent with Muse Spark 1.2
What happened: Meta released Muse Code, a terminal-based coding agent in beta for macOS and Linux, powered by the new Muse Spark 1.2 model. It keeps background agents active across long sessions and logs every action for crash recovery.
Why it matters: Long-horizon coding agents that retain context and survive crashes could handle multi-hour tasks, reducing developer oversight.
Source: The AI & Software Report
Alibaba Launches Wan3.0 Video Model with 30-Second Generation
What happened: Alibaba Cloud launched Wan3.0, a video-generation model that creates clips up to 30 seconds and accepts DOC, XLS, PPT, PDF, and Markdown as inputs. It had been in public beta since early August.
Why it matters: Document-to-video generation opens new workflows for developers building automated content pipelines.
Thomson Reuters Launches Proprietary LLM ‘Thomson’
What happened: Thomson Reuters announced Thomson, its first proprietary large language model, trained in-house for $40 million from an open-source foundation. The company says it runs at a fraction of the cost of comparable frontier models.
Why it matters: A domain-focused model built on proprietary data at low cost could challenge general-purpose frontier models in legal and professional services.
Nvidia Unveils Vera Rubin LPX and CPX Inference Platforms
What happened: Nvidia announced the Vera Rubin rack-scale system with Groq 3 LPX in full production, delivering 3, 400 output tokens per second on Gemma 4 31B in an Artificial Analysis benchmark, 4x faster than the nearest alternative. CoreWeave and Nebius are adopting the platforms.
Why it matters: Faster, cheaper inference for long-context agentic workloads could lower token costs and enable more complex AI applications.
Japan Drafts IP Code for Generative AI Training Data
What happened: Japan’s Cabinet Office proposed a revised Principle-Code for generative AI businesses, covering IP protection, avoiding pirate sites, respecting paywalls, and increasing transparency. It uses a ‘comply or explain’ approach and applies to foreign businesses serving Japan.
Why it matters: Builders targeting Japan may need to document training data practices and IP risk management, even if non-binding.