Models

Alibaba launched Happy Horse Friday and watched it climb to the top of the global text to video rankings within hours. Shares rose 2.12 percent on the announcement. It is the second Chinese lab this week to land a model that the Western frontier cannot ignore.

A one trillion parameter Mixture of Experts model trained for roughly $5.2 million on Huawei Ascend 950PR chips. One million token context window, Apache 2.0 license, and no Nvidia hardware anywhere in the pipeline.

Meta Superintelligence Labs, the division Alexandr Wang was hired to build, unveiled Muse Spark Wednesday. It is natively multimodal, tool using, multi agent orchestrating, and closed. The company that built the case for open weights just folded it.

Nine days after GPT 5.4, OpenAI added a one hundred dollar Pro tier with unlimited GPT 5.4 access and ten times the Codex usage of Plus. It also shipped GPT 5.3 Instant Mini as a new fallback model. The pricing ladder is forming in public.

A new general purpose frontier model restricted from public release. Anthropic says Mythos has already flagged thousands of zero day issues across major operating systems and browsers, and the Glasswing consortium gives twelve large companies defender first access.

GPT 5.4 scores 75 percent on the OSWorld V benchmark, ships with a context window of one million tokens, and lands alongside a redesigned ChatGPT that collapses chat, code, search, and agents into one product for 900 million weekly users.

Four open weight models built for advanced reasoning and agentic workflows. The 31B variant is the third ranked open model globally on the Arena AI text leaderboard, shipping under a permissive license.