Cur8 — Wednesday, September 02, 2026

Quasar 438B scores 43 on the Artificial Analysis Intelligence Index, outperforming European models like Mistral Medium 3.5 (30) and Nemotron 3 Ultra (38), while answering 500 tokens in 15.3 seconds, faster than most higher-scoring models.

BenchMIRT identifies hidden capabilities in LLM benchmarks by analyzing individual questions, revealing that some safety benchmarks like BBQ and WMDP correlate more with general reasoning than safety. Trained on 100 LLMs and 34K questions, it disentangles mixed signals.

Top open-source coding models in 2026 include GLM 5.3 (1M context, 743B params), Kimi K 2.7 (1T params, agentic tasks), DeepSeek V4, Qwen 3 Code (180B params), and Qwen K3 (2.8T params). For local use, Qwen 3.6 27B (17GB) and Codestral 2 (FIM training) are key choices based on hardware and task needs.

Users can opt out via admin panel or mobile app settings.

Astra meets OpenAI's Critical cybersecurity capability threshold, featuring enhanced safeguards for secure deployment.

Apple presents forensic evidence from ex-employee Chang Liu's MacBook, showing he used stolen Apple schematics at OpenAI and attempted to destroy evidence. Apple seeks expedited discovery, claiming trade secrets are being used.

AI-native companies use agents to automate onboarding, account management, and developer tools, improving operational efficiency for enterprises.

Healthcare organizations can now link EHR and industry data to ChatGPT, enabling clinicians to securely access patient context and medical research.

Gilbert + Tobin scales ChatGPT Enterprise and Codex with CEO-led governance, ensuring human accountability and rigorous oversight across AI deployment.

Claude Fable 5.1 and Mythos 5.1 offer 25-45% lower pricing, improved safeguards, and advanced capabilities in coding, scientific research, and biology, with Mythos 5.1 achieving 2.5x speedups in computational biology.

Claude Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, with max effort producing detailed 65,927-token output costing $3.30.

Claude 5.1 now refuses to reproduce song lyrics, copyrighted visuals, or recognizable characters, with persistent refusal after initial decline. It also provides harm-reduction info for drugs without dosing guidance, and avoids apologizing to rude users.

Claude Fable 5.1 improves on Fable 5 with better reasoning, agentic coding, and end-to-end knowledge work, available on AWS with 30-day data retention and optional zero data retention for EFS customers until 2026.

Google AI advances in August 2026 include Gemini 3.7 Flash at half the price, Pixel 11 with Gemini Nano, and 1 billion Gemini app users.

Google Pics, based on Nano Banana, rolls out to Workspace users, enabling image creation and editing within Docs, Slides, and Drive. Features include object segmentation, in-image text editing, and multiple generations from a single prompt.

DeepMind boosts accuracy by 7% in video analysis, enabling sub-second retrieval and anomaly detection. The feature cuts video analysis costs by up to 66%.

Inference engineering optimizes LLM deployment by managing tradeoffs (e.g., latency vs. throughput) and pushing efficiency frontiers. Techniques like batch sizing, parallelism, quantization, kernel optimization, speculative decoding, and disaggregation improve performance, with some enabling targeted tradeoffs and others expanding overall efficiency.

New NYC policy prohibits AI use by students through 8th grade, effective immediately, aiming to prevent academic misconduct and ensure equitable learning.