xAI Releases Grok 4.3 — Aggressive Pricing and Imagine Agent Mode
xAI has released its new model, Grok 4.3. Its biggest highlights are a major price cut and improved tool-use capabilities. While the model shows better performance on practical tasks, it is still considered a step behind the top models from OpenAI and Anthropic.
The company has also introduced a new agentic image generation feature called "Imagine agent mode." Designed for creative projects, it generates and iterates on images while interpreting the user's intent.
xAI is clearly betting on low prices to expand its market share, and price competition in the LLM market looks set to intensify even further.
OpenAI Enables Ad Tracking by Default for Free ChatGPT Users
In countries where it serves ads, OpenAI has enabled marketing cookies by default for free-tier ChatGPT users. Paid subscribers are exempt, and users can opt out through the settings.
With ChatGPT's advertising model now in full swing, it has become clear that OpenAI's monetization strategy is shifting toward a twin-pillar model of "subscriptions + advertising." Privacy concerns have also surfaced over free users' data being used for ad targeting.
Unsloth Identifies and Fixes Inference Bug in Mistral Medium 3.5
Working with Mistral, the Unsloth team has fixed a bug that was affecting inference in Mistral Medium 3.5. The cause was a parsing issue in YaRN (an extension of Rotary Position Embedding), resolved by changing the `mscale_all_dim` setting from 1 to 0.
The issue affected multiple implementations, including transformers and llama.cpp, and in particular caused severe degradation in output quality on long contexts. Fixed GGUF files are already available on Hugging Face.
Reports are pouring in from the local LLM community that the fixed model is now far more stable even on short contexts.
AI Silicon in 2026 — Custom Silicon Won't "Replace" NVIDIA
Hyperscalers are rolling out custom AI chips one after another: Google TPU v7 (Ironwood), AWS Trainium2/3, Microsoft Maia 100/200, and Meta MTIA v3. However, an analysis gaining traction on Qiita argues that the predictions from around 2024 that "custom silicon will dethrone NVIDIA" are, in reality, heading toward a "division of labor."
Each company's in-house chips are optimized for specific workloads, but NVIDIA's strengths — its CUDA ecosystem and versatility — remain unshaken. Rather than a binary "NVIDIA vs. custom silicon" standoff, the AI silicon market in 2026 appears to be evolving toward the optimal combination of chips for each use case.