Cur8 — Saturday, July 25, 2026

Claude Opus 5 leads AI leaderboard, matches Fable 5's intelligence at half the price, excels in proactive tasks, and improves cybersecurity vulnerability detection. Priced same as Opus 4.8, with fast mode option.

Opus 5 excels in prompt injection resistance, outperforming predecessors in evaluations.

Claude Opus 5 tops AI leaderboard with score of 61. Mercury 2 fastest at 901.6 tokens/s, Nova Micro cheapest at $0.03/M tokens.

Gemma 4B outperformed Llama 3 1 8B and Qwen 2 5 7B in local testing, excelling in speed, accuracy, and ease of use on a Mac Mini with 16GB RAM. Gemma was the fastest, consistently correct, and required the least effort to get useful answers.

OpenAI's rogue agent story, similar to GPT-2's release, aims to attract investment and regulatory favor by emphasizing AI's power and danger. Critical analysis reveals AI's cybersecurity capabilities can benefit both attackers and defenders, but current US restrictions on model access may hinder defensive efforts.

U.S. tech giants warn against restricting open-weight AI models, citing competition and innovation benefits. Chinese models like Moonshot AI's Kimi K3 outperform U.S. offerings in some benchmarks, sparking debate on potential restrictions.