Which AI writes the better take? You decide — blind.

Top Chinese AI labs are already using Nvidia’s Blackwell GPUs for their newest model builds. For anyone shipping LLM products, the practical takeaway is that Chinese model competition should be treated as Blackwell-class, not constrained-to-older-hardware; expect stronger open and commercial models, faster benchmark movement, and less room to rely on hardware access as a durable moat.

Gemini 4 Hydrogen is only 25% cheaper than Carbon/Argon today, won’t become a 75% discount until December 1, and is still 15x the $0.10/$0.50 pricing of Claude Haiku 5.5 and GPT-6 Luna. For production routing, Hydrogen looks less like a broadly cheap default and more like Google’s free-tier/Search and multi-model execution model: useful for keeping Gemini-family workflows on-platform, but hard to justify on cost alone if you can route routine tasks to cheaper competitors.

Granite 4.1 models now support LLMON, a markup language that explicitly communicates structure and semantics to LLMs using reserved special-token delimiters, reducing brittleness and improving robustness against prompt injection attacks; this enables more reliable processing of complex context with multiple components, such as instructions and data payloads, in production LLM applications.

Amplitude reduced product improvement shipping time to 1-2 days by leveraging OpenAI models to power Wave, an AI agent that identifies, implements, and measures product changes, enabling rapid iteration and potentially disrupting traditional development timelines for teams shipping similar products.

Gemini's new Decision 1 model supports multimodal input including video, audio, and PDFs, at a cost of 10 cents per million input tokens. This enables developers to integrate multimodal decision-making into their production LLMs and agents, expanding their capabilities beyond text-only inputs. It directly competes with OpenAI's decision model, offering a potentially different trade-off in terms of confidence scoring and cost.