AI Weekly Recap: Ads in ChatGPT, Anthropic's $30B Round, and the SaaS Trillion-Dollar Dip

What a week. OpenAI started testing ads in ChatGPT, Anthropic countered with a $30 billion funding round and a clear "Claude stays ad-free" stance, and somehow Claude ended up driving a rover on Mars. Here's everything worth knowing.

OpenAI Tests Ads in ChatGPT

OpenAI is officially testing advertisements inside ChatGPT, with privacy controls for users. It's a predictable move given the economics of running frontier models at consumer scale, but it's already drawing criticism. The question is how long the ads stay "subtle."

In other OpenAI news: Deep Research now runs on GPT-5.2 and reportedly found a new result in theoretical physics. GPT-5.3-Codex-Spark launched as a real-time coding model with 15x speed for Pro users, running on a dedicated custom chip. And the legacy spring cleaning has begun — GPT-5, GPT-4, and several variants are being shut down.

OpenAI also accused DeepSeek of distilling US AI models for R1 training, launched a secure ChatGPT version for the US military, and the Jony Ive-designed hardware device got pushed to 2027 due to a trademark lawsuit.

Anthropic Raises $30 Billion at $380B Valuation

Anthropic closed a $30 billion Series G round with a post-money valuation of $380 billion. Not bad for a company that's barely three years old.

The Super Bowl ad paid off — Claude shot into the App Store top 10 right after. The company also launched self-serve enterprise plans so organizations can sign up directly without a sales team, and shipped Cowork on Windows with full feature parity to macOS.

On the people side, a WSJ profile highlighted Amanda Askell, the philosopher shaping Claude's ethics. Meanwhile, safety researcher Mrinank Sharma resigned with a public warning about global AI risks. When the safety researcher leaves because things feel too risky, that's worth paying attention to.

New Models: Claude Opus 4.6 and GPT-5.3 Codex

Claude Opus 4.6 and GPT-5.3 Codex dropped within 20 minutes of each other. Both are extremely strong at coding, but the real story is the capability overhang — these models can already do far more than most people extract from them. Power users leverage 7x more reasoning and analysis capabilities than average users.

Opus 4.6 brings a 1-million-token context window (roughly 1,500 pages of text) and adaptive thinking with effort control. GPT-5.3-Codex is about 25% faster for agentic workflows and reportedly helped build itself.

The action is increasingly happening outside the chat window: better workflows, tool integration, autonomous agents.

The SaaS Trillion-Dollar Dip

The software sector lost a trillion dollars in market cap. Asana, DocuSign, ServiceNow, Salesforce, Adobe, Figma — all in freefall. Even Microsoft dropped 5%.

The trigger? Anthropic's Cowork plugins demonstrated that AI agents can replicate specialized software workflows end-to-end. Productivity, project management, marketing, legal, finance, data, enterprise search, customer support — all covered as continuous workflows instead of tool silos.

The market's fear: if AI models handle standard workflows, why pay for SaaS subscriptions? Not every SaaS tool will disappear, but many will need to redefine their reason for existing. This isn't a random dip — it's a structural shift.

Super Bowl Goes All-In on AI

15 of 66 Super Bowl ads came from AI companies — nearly one in four. Anthropic, OpenAI, Google, Amazon, Meta, plus startups like Genspark, Ramp, and Wix. Thirty seconds of airtime costs around $10 million. AI has officially gone mainstream.

But there's a déjà vu here. Four years ago it was crypto. FTX, Crypto.com, Coinbase, eToro — all had Super Bowl spots. Nine months later, FTX collapsed and $32 billion in valuation evaporated.

AI has more substance than crypto did — real use cases, real revenue, hundreds of millions of active users. But there's also a lot of hype. Stay skeptical, use the tools that actually work, and ignore the noise.

Perplexity Launches Model Council

Perplexity shipped Model Council, and it's exactly what multi-model workflows needed. You ask a question, it goes to three models in parallel (say, Claude Opus 4.6, GPT-5.2, and Gemini 3.0), a synthesizer model compares the answers, and you get a response with clear markers showing where models agree and where they disagree.

All three agree? Go with it. Contradictions? Now you know where to dig deeper. Every AI model has blind spots. MIT researchers tested math problems — a single model hit about 70% accuracy. Three models debating? 95%. That's the difference between useful and reliable.

ByteDance Drops Seedance 2.0

ByteDance released Seedance 2.0 in beta, and early tests look impressive. With access to millions of videos for training, the results for UGC content are strong.

The model handles multi-shot storytelling (multiple scenes, same character, consistent style), generates native audio matched to video, delivers 2K resolution with 15-second output, and produces usable results on over 90% of first attempts.

It's currently only available in China on Jimeng AI, but expansion to other platforms is expected by the end of February. It immediately drew pushback though — SAG-AFTRA accused ByteDance of using members' voices and likenesses without permission, and the Motion Picture Association called for it to be shut down over copyright infringement.

Quick Hits

• Google/DeepMind upgraded Gemini 3 Deep Think, their specialized reasoning mode for math, physics, and computer science.
• MiniMax launched M2.5 with state-of-the-art productivity and lower operating costs.
• Alibaba introduced Qwen-Image-2.0 for advanced image generation.
• Runway shipped Story Panels for AI-generated film and content production.
• Meta is reportedly building facial recognition into smart glasses, with a potential launch this year.
• Spotify says their best developers haven't written a line of code since December, thanks to AI coding tools.
• xAI lost two co-founders in two days. Not a great look.
• T-Mobile brought live phone call translation — even to old flip phones.

That's the week. The pace isn't slowing down.