Curated AI news from the last 24 hours — models, agents, enterprise, and developer tools.
xAI officially launched Grok 4.6, its flagship model for coding, agentic tasks, and knowledge work, with a 500,000-token context window and configurable reasoning levels for long-running agents. The model was subsequently added to Google's Enterprise Agent Platform Model Garden, expanding its availability for enterprise agentic and visual-work use cases.
docs.x.ai ↗OpenAI released early benchmark results for Jalapeño, its first custom inference chip built with Broadcom. On the SemiAnalysis InferenceX benchmark, Jalapeño-based systems delivered 1.5x-1.9x more throughput and 1.7x-3.6x lower latency than current state-of-the-art GPU systems across models like GPT-OSS-120B and DeepSeek R1, while using significantly less power per rack.
techcrunch.com ↗Cursor launched Origin, a Git-hosting platform built into its AI coding editor that offers pull-request review, code search, and two-way sync back to GitHub. The launch marks Cursor's shift from a tool that sits on top of GitHub to a direct competitor, built in part on technology from its December 2025 acquisition of code-review startup Graphite.
techcrunch.com ↗Apple unveiled the M6 chip, its first built on a 2-nanometer process, debuting in the new Mac mini with a dual 16-core Neural Engine delivering up to 2x the prior generation's peak AI compute. The M5 Ultra, Apple's first quad-die chip, powers the new Mac Studio with up to an 80-core GPU and 512GB unified memory, delivering up to 4.5x the peak GPU AI compute of the M3 Ultra for running large models locally.
apple.com ↗Apple is building waitlist infrastructure into iOS 27 for its rebuilt, generative AI-powered Siri, meaning not all users will get immediate access at launch this fall. The system will show dynamic wait-time messages as users are gradually let in, similar to staged rollouts Apple has used for other high-demand features.
macrumors.com ↗Ahead of a planned October IPO targeting a roughly $2 trillion valuation, Anthropic is pitching investors on a total addressable market of $30 trillion, nearly equal to annual U.S. GDP. Analysts note the company's current revenue run rate is far below what such a valuation would require at typical tech multiples, raising scrutiny over the underlying math.
fortune.com ↗Anthropic launched a $5 million grant program funding independent research into how AI impacts user wellbeing, providing recipients with direct funding, model access, and technical support. The program targets open-source evaluations of how conversational AI should behave when users seek emotional support, requiring funded work to involve clinical experts and reflect real multi-turn conversations.
anthropic.com ↗OpenAI published data showing enterprise AI usage is shifting from assistance to execution of full workflows, with the top 10% of firms by AI usage generating 8.3x as many output tokens per active user as typical firms. The report highlights a widening gap between frontier adopters and the rest of the market as companies move AI agents into production.
openai.com ↗Amazon is shutting down Mechanical Turk, its 21-year-old crowdsourced human-labeling marketplace, effective September 30. The company pointed to the platform's decline as AI models have reduced reliance on human data annotation and newer entrants like Scale AI, Mercor, and Prolific have taken over the data-labeling market.
cnbc.com ↗Z.ai, maker of the GLM model family, confirmed it created Ox Alpha, an open-weight model that had topped several benchmark leaderboards under a mysterious identity. The company said it plans to release the model's weights publicly, continuing the trend of Chinese labs releasing competitive open models.
techcrunch.com ↗IBM released Granite 4.2, an Apache 2.0-licensed family of open-weight models in 3B, 8B, and 30B sizes, pre-trained on 15 trillion tokens with a 512,000-token context window. The larger 8B and 30B models went through additional agentic reinforcement learning to call tools, edit and run code, and search the web, and can toggle between thinking and low-effort reasoning modes.
thenewstack.io ↗