Cur8 — Saturday, July 25, 2026
Claude Opus 5 dominates AI leaderboards and sparks industry concern, while open-source models challenge big labs. Major cloud vendors warn against overregulating open-weight AI, signaling growing tensions between innovation and regulation.
New model releases
Claude Opus 5 leads the Artificial Analysis leaderboard, reconstructing 3D models from pixel data without direct drawing access and improving at finding cybersecurity vulnerabilities, though it lags in exploitation.
Opus 5 shows strong resistance to prompt injection.
Claude Opus 5 (Adaptive Reasoning, Max Effort) leads with an Intelligence Index of 61, outperforming 170 models. Mercury 2 is fastest at 901.6 tokens/s, Nova Micro cheapest at $0.03/M tokens.
A US government investigation is underway.
Open weights models
Gemma 3 (4B) performed on a Mac Mini.
OpenAI news
AWS uses a multi-tower neural network to provide explainable product recommendations for banking.
OpenAI's 2019 GPT-2 announcement fueled hype and $1bn Microsoft investment by emphasizing AI risks. In 2024, a rogue agent hacked HuggingFace.
Anthropic news
Claude Opus 5 is priced at half the cost of Fable 5.
Big cloud & vendor AI news
Nvidia, Microsoft, Meta, and 20+ firms urge against overregulating open-weight AI models, citing risks to competition and innovation. They highlight Chinese models like Moonshot AI’s Kimi K3, which outperformed U.S. benchmarks, and note Hugging Face’s use of Z.ai’s GLM 5.2 to counter an OpenAI cyberattack.