Models & Research

OpenAI shipped an Admin plugin for ChatGPT Work and Codex on August 25, letting admins read usage data and change workspace settings inside one conversation.

OpenAI published the first InferenceX benchmark results for Jalapeño, its custom inference chip, claiming 1.5 to 1.9 times more AI work per watt than the Nvidia GB200 and GB300 systems it was measured against.

IBM released Granite 4.2 on August 25, 2026 in 3B, 8B and 30B sizes under Apache 2.0. Only the 8B and 30B got the agentic reinforcement learning stage.

Sentence Transformers v6.0 adds a MultiVectorEncoder class for ColBERT-style late interaction retrieval. A 14.5-hour finetune on one RTX 3090 beat every zero-shot retriever Hugging Face tested on medical search.

Z.ai says the anonymous Ox Alpha model was GLM-5.3-Flash, now on Hugging Face under an MIT licence, and that it served the whole stealth week on Chinese AI chips.

Apple refreshed the Mac mini and Mac Studio on August 25 around the M6 and M5 Ultra, with prices from $899 to $5,499 and a 512GB memory ceiling aimed at local AI inference.

AWS switched on cross-Region inference for GPT-5.6 on Amazon Bedrock in more than 25 Regions. Global routing widens the capacity pool, but output tokens still burn quota at 10x.

Gruber says watermarking makes Claude pick words for detectability, not clarity. Anthropic says nothing is lost. Coders shipped removal tools within four hours.

Google launched AI study tools across Search and Gemini on August 19, 2026, with practice quizzes for nine standardized exams and a free year of a paid AI plan for eligible college students.

AWS published ADOP, a reference architecture where Claude Code sub-agents on Amazon Bedrock generate ETL, quality checks and Cedar policies at build time, so production runs deterministic code with no model in the loop.

OpenAI’s new plugin lets ChatGPT read, search and send Apple Messages on the Mac, in exchange for Full Disk Access and a stack of permissions.

Together AI ran 904 DeepSWE rollouts: GLM-5.3 tied Claude Fable 5 on first-attempt accuracy at $3.99 a rollout against $21.63, and a GLM-first cascade beat GPT-5.6 Sol.