Good evening, everyone, and welcome to today's AI Evening News. At noon today, Google employee Logan Copatrick posted a message containing only the word "GEMINI," which the community took as a signal of a new model release. Third-party leaks followed, claiming that Google's unreleased Gemini for Hydrogen can already be invoked on Vertex AI. One account said it had already appeared in the SKU and billing catalog, with pricing possibly at a discounted rate: $0.75 for input and $3.75 for output. According to reports, Google had previously let employees test it on an internal platform, with one employee saying it resembles the Flash series—faster and cheaper than A-Gang.
Google has not yet officially confirmed any of the above. Meanwhile, according to reports, three of OpenAI's national security affairs leads were tricked into a video call last month by a Russian prank duo posing as the head of Ukraine's foreign intelligence service, during which the two sides discussed providing military AI support to Ukraine. OpenAI said it has launched an investigation and confirmed it was a prank. Developers can insert an empty system message and specify a new effort level, allowing them to adjust the effort level mid-conversation—changing how hard the model works on its answers—without affecting prompt caching. Google has released version 1.0, the first stable version of its generative UI open protocol A2UI, which lets AI agents respond with interactive interfaces such as forms and charts; web and mobile rendering SDKs are now available. Volcano Engine has released the video generation model Sedence 2.5 Flash, whose quality approaches that of the flagship model while generating twice as fast. 720P video generation is priced at about 0.99 yuan per second, roughly 34% cheaper than the flagship. The model will roll out on platforms such as Jimeng starting today, with the API launching on November 15. Volcano Ark's Doubao workspace announced product updates: in addition to its own Doubao 2.1 Light and Pro, it has integrated third-party models such as Kimi K3.1 Pro and GLM 5.3, and launched a limited-time promotion lowering the quota consumption multipliers for some models. New users who log in on the desktop version can claim a free 30-day subscription. According to reports, ByteDance founder Zhang Yiming recently said at an internal meeting with AI teams that ByteDance has the capability to compete with OpenAI and Anthropic, but truly catching up with frontier technology would take until at least 2029 or 2030. He also proposed that ByteDance's large models should benchmark against the international first tier. Baidu announced the launch of a large model token trading platform called Resource, which claims to be neutral and trustworthy, testing models upon onboarding, continuously monitoring quality, and assigning credit ratings to model providers. OpenRouter announced that its website has launched beta localization in seven languages, including Simplified Chinese and Japanese. Its translation system used GPT 6.1 SL to complete the translations, with automatic error correction through code checks and model verification. That's all for today's news—see you tomorrow.