AIThe openai/math repository groups the papers into 372 families. The model was posed about 4,000 problems; the average result used roughly three hours of ChatGPT Pro thinking compute.
AIThe preview API is live; open weights follow by the end of October. Mistral says the model was trained from scratch in its own European datacenters and leads open models on several cyber tests.
AIIn response to the EU AI Act, ChatGPT and Codex output in the EU gets an invisible watermark; API customers can opt in, and the detector stays closed at launch.
AIDevice data is decrypted only inside trusted execution environments running published training code, and the access policies sit in a public transparency log that outsiders can audit.
AIAcross 507 business workflows, two thirds of failed attempts ended cleanly with no tool error. Kimi-K3 solves the most tasks once; Claude Opus 5 and 5.5 pass the most every time.
AIFine-tuned from Qwen3-8B on 47,000 filtered reports, it cuts Asta's report time to about 51 seconds from 178. Weights and training data are public; the evaluation dates from 2025.
AIA teacher model shows what success looks like, a generator writes new tasks around the gap, and two gates reject tasks a verifier would pass wrongly. Pass@1 rose 7.2 points.
AIThe open framework moves mixture-of-experts training from FSDP to DDP with experts kept on the GPU. Growing the expert pool from 8 to 128 cost under 5% throughput.
AIError rate, speed and voice similarity replace weeks of arena voting. Kokoro-82M, supertonic-3 and s2-pro lead on English; OmniVoice is among the multilingual leaders.
AIThe upgrade matches Astra on the DeepSWE coding benchmark at roughly one-fifth of the cost, and cuts cached input to $0.10 per million tokens. Astra stays ahead on the hardest science tasks.