Major Model Updates

• Google's Gemini 3 Deep Think: A major upgrade expands its reasoning capabilities into science, engineering, and research. The model sets new industry benchmarks, scoring 84.6% on ARC-AGI-2 and 48.4% on Humanity's Last Exam, while achieving gold-medal results at the 2025 International Physics and Chemistry Olympiads. The update is available for Google AI Ultra subscribers.

• OpenAI's GPT-5.3-Codex-Spark: A new, faster coding model that delivers over 1,000 tokens per second. It is OpenAI's first product to run on chips from Cerebras, an Nvidia rival, marking a milestone in a $10B+ partnership. It is available as a research preview for ChatGPT Pro users.

• ByteDance's Seedance 2.0: This new AI video model from the maker of TikTok is gaining widespread attention for its hyper-realistic output, significantly reducing the distortions and unnatural movements common in previous models.

Industry, Funding & Geopolitics

• Anthropic's Landmark Funding: The company raised $30 billion in a Series G round at a $380 billion valuation. Anthropic also announced its revenue run rate has hit $14 billion and donated $20 million to a PAC supporting AI regulation.

• AI in the Workplace: Spotify reports its top developers haven't written code since December 2025, instead using an internal AI system called 'Honk' (built on Claude Code) to handle bug fixes and feature development.

• Model Cloning & Security: OpenAI and Google have warned lawmakers about Chinese firms like DeepSeek using "distillation attacks" to clone their models. This raises concerns about competitor models being created without necessary safety guardrails. The Pentagon is also pushing AI labs to deploy models on classified military networks.

The Open-Source Surge from China

• Chinese AI labs are showing increasing innovation in open-weight models, creating a competitive alternative to US-based closed models.

• MiniMax launched M2.5, an open-source model that rivals the coding performance of GPT-5 and Claude Opus 4.6 at a fraction of the cost, making it viable for long-running AI agents.

• Zhipu AI released GLM-5, an open-source model specifically designed for agentic tasks.

AI Benchmarking & Research

• Major AI labs are now using the video game Pokémon as a complex benchmark to test AI agent capabilities.

• The game requires sustained planning, resource management, and error recovery over hundreds of hours, simulating real-world agentic challenges better than games like chess or Go.

• Google's Gemini was the first to finish Pokémon Blue, completing it in approximately 406 hours.