Good morning. Capital is the dominant theme today: OpenAI is reportedly chasing a $1.25 trillion valuation, DeepSeek pulled in $15 billion two days after being kicked off US federal systems, and the top AI labs just hit record lobbying quarters in Washington. Meanwhile Microsoft is pitching “medical superintelligence,” and MIT researchers found that AI agents left alone will apparently invent games.
OpenAI eyes $1.25T, DeepSeek grabs $15B. OpenAI is in advanced talks for a $30 billion raise at a $1.25 trillion pre-money valuation led by Coatue and UAE investors, which would make it the most valuable private company on record and comes with a side deal for additional compute from the UAE, TechCrunch reports. The round precedes a planned IPO for which OpenAI filed confidentially in June, and Anthropic is reportedly circling a public debut of its own at $1.5-2T. On the other side of the restriction wall, DeepSeek raised roughly $15 billion from backers including CATL and Tencent — one of China’s largest-ever rounds — arriving 48 hours after DHS ordered federal agencies to remove its products.
Record lobbying quarters in DC. OpenAI, Anthropic, and Nvidia each posted record Q3 2026 lobbying spends totaling nearly $6 million, with Nvidia at $2.2M, Anthropic at $2.1M, and OpenAI at $1.5M, The Verge reports. The issues range from child safety and export controls to data center energy costs and Trump’s executive order rebranding AI as “Super Intelligence.” Midterms are coming, and the labs are visibly bracing for a more regulation-friendly Congress.
Microsoft claims its AI beats doctors on hard cases. Mustafa Suleyman is calling Microsoft’s new MAI-DxO diagnostic system a step toward “medical superintelligence,” citing a Science paper where the multi-agent tool hit 83% accuracy on complex cases versus 36% for generalist physicians. The caveat, as TechCrunch points out: complex cases in the real world go to specialists or care teams, neither of which were in the benchmark. Most of the gain also seems to come from the underlying GPT-5 and GPT-5.5 models rather than Microsoft’s orchestration layer, and improvements on emergency cases were small.
MIT’s agents invented their own games. Researchers dropped 13 Claude Opus 4.7 agents onto a simulated desert island for 30 hours with no instructions and watched them spontaneously generate goals and challenges — behavior the team is comfortable calling play, TechCrunch reports. Agents given this unstructured time later outperformed identical agents without it on specific tasks, hinting at goal-free exploration as a training paradigm (particularly for robotics). Six of the 13 also independently discovered a physics exploit letting them throw objects farther by releasing mid-jump, which is either charming or ominous depending on your priors.
1Password wants to broker agent logins. 1Password launched 1Password for Agents, which issues agents a unique digital key, pings the user for confirmation, and fills credentials itself so agents from Anthropic, OpenAI, Meta, and xAI never actually see the password. The timing is pointed: US banks are reportedly detecting AI agents operating inside customer accounts with no way to tell if the activity is authorized, with one banker calling the situation “horrifying.” Credential infrastructure is scrambling to catch up with agent adoption.
Canva ships its own image model, Anthropic doubles down on Duet. Canva released Canva Craft, a proprietary two-stage image model that plans spatial layout and text before generating, trained on millions of human-made templates and producing editable layered designs up to 2K. It currently sits fourth on Design Arena, behind three OpenAI GPT-Image models, and brings generation in-house to cut API costs. Anthropic, meanwhile, opened Claude Duet more broadly to Pro and Max users on iOS, Android, and web — the Opus 5.5 agent with 6,000+ connectors we flagged yesterday, now with Stripe and Shopify purchase support.
That’s a lot of money moving in one Wednesday. Watch whether OpenAI’s UAE compute arrangement draws the same export-control scrutiny DeepSeek just did.