A new AI model named OX Alpha has stormed onto the scene, boasting a staggering one million token context window and multimodal capabilities. At the same time, separate leaks reveal a powerful GLM 5.3 Flash model and Tencent’s HY4—all pointing to a fierce race in Chinese AI development.
OX Alpha Shakes Up the AI World
OX Alpha emerged seemingly out of nowhere, debuting on platforms like Open Code and Open Router with features that sound like science fiction. This mysterious AI boasts a 1 million token context window, multimodal abilities, and the infrastructure to handle up to 100 trillion tokens daily. What’s astonishing is that it’s accessible for free, allowing early testers to push its limits and quickly notice its advanced reasoning and visual capabilities.
Though the model points to being Chinese, no one knows its exact origins. Xiaomi and GLM are the frontrunners based on behavioral patterns and token quirks, but no confirmation exists. The model even rendered a black hole simulation, showing light bending in a way that’s both eerie and impressive. Its ability to visualize complex 3D environments extends to creating a Minecraft-like sandbox game featuring mobs and infinite terrain, an indicator of substantial coding prowess.
Tests also show OX Alpha simulating Cerebras’ CS4 hardware in 3D, rivalling the output seen from Claude Opus 5. Taking all evidence together, many experts lean towards it being a new version of the GLM model family, though the secret remains firmly under wraps.
GLM 5.3 Flash: Multimodal and Ready for Expansion
Adding to the hype, a totally separate leak reveals the existence of the GLM 5.3 Flash model, internally tested and packed with exciting upgrades. Unlike previous GLM iterations, this one incorporates a video encoder, a tokenizer reminiscent of its predecessors, similar response styles, and crucially, multimodal capabilities.
This leap means GLM is catching up to user demands for vision and audio integration, expanding its reach beyond pure text generation. Instead of retraining a base model, GLM’s team used large-scale post-training and reinforcement learning, yielding major improvements in coding and handling lengthy agentic reasoning tasks.
Tencent Joins the Race with HY4
China’s AI boom isn’t stopping there. Tencent’s HY4, spotted in a gray testing phase within its app, is marked as an expert-level model surpassing both HY3 and DeepSeek. The company has openly committed to releasing HY4 later this year, claiming it’s on par with top AI contenders like Opus 5.
This means three separate high-profile Chinese AI models—OX Alpha, GLM 5.3 Flash, and HY4—are creating waves simultaneously, underscoring how heated and fast-moving the AI development scene has become in China.
OpenAI’s Astra Delayed as Safety Takes Priority
Turning to global giants, OpenAI’s Astra release faces another delay as the team squares away alignment issues and clamps down on reward hacking, a behaviour that lets AI exploit its training goals unfairly. Instead of launching a new model, Astra has received an update to refine these aspects before going public.
Anthropic is reportedly holding back Fable 5.1 to coincide with Astra’s eventual release, hinting at a tech showdown coming soon. Expect Astra’s launch sometime in September now, pushing back what was initially anticipated.
Controversies Over Codex Usage Limits
Meanwhile, OpenAI users report abrupt drops in Codex usage limits, with some pro users burning through their allocations far quicker than usual. After investigations, OpenAI denied secretly lowering limits but linked issues to the use of third-party API subscription models. This has led to frustration within developer communities who rely on Codex for coding assistance, casting a shadow on OpenAI’s service transparency.
Anthropic Adjusts Enterprise Data Policies
Competing in data privacy, Anthropic announced plans allowing enterprise clients to control their safety data themselves rather than letting the company store sensitive prompts and outputs. This follows OpenAI’s recent private safety processing efforts targeting the same problem—balancing AI oversight with client confidentiality.
New Image Model and ChatGPT Features Emerge
OpenAI is quietly testing a new image generation model codenamed Mona Lisa 1, which recent evidence suggests might be an upgraded GPT image model, possibly version 2.5. This development hints at continued refinement in AI’s visual creativity, alongside a small update introducing an Apple Messages plugin for Mac users to better manage conversations within ChatGPT workflows.
Anthropic recently launched Claude Academy, a free platform to help users improve their Claude AI mastery—from beginners to seasoned pros.
What’s Next in AI?
The sudden floods of new models, leaks, and experimental releases reflect the intense competition shaping the AI landscape. The rise of Chinese labs alongside Western powerhouses is accelerating innovation and expanding capabilities, from extended context windows to true multimodal understanding.
Anyone deeply invested in AI’s future should keep an eye on these developments. With models like OX Alpha and GLM 5.3 Flash introducing features many have long awaited, the next phase of AI might look and act very differently from today’s familiar tools.
Rafomac News, Tech & Trends That Matter