Thinking Machines, founded by Mira Murati, has released its first open-weight model Inkling, based on a MoE architecture with approximately 975B total parameters and 41B active parameters. The model natively supports text, image, and audio reasoning with a context window of up to 1 million tokens, and is released under the Apache 2.0 license. It is available on Hugging Face and can be fine-tuned on Tinker. The model emphasizes customizable fine-tuning, reflecting the team's engineering-first philosophy that 「models are not the endpoint—customization is」. Multiple evaluations show Inkling performs comparably to GPT-5.6 Sol on design arena, positioning it as a top-tier open-source model outside China, between Kimi 2.5 and 2.6. A smaller 12B variant is also in development. Fireworks AI already offers day-zero deployment support, and Inkling has entered LMArena's Agent Arena.

OpenAI has unveiled its first consumer-grade AI hardware—a portable, screenless smart speaker designed to serve as a home computer and human-like AI companion in the AI era. The device enables smart home control, media playback, and question answering, emphasizing personalized interactions. Previously, Apple filed a lawsuit against OpenAI over trade secrets, which may impact its hardware development plans. Meanwhile, Apple is in talks with startup PrismML to compress large AI models for direct execution on iPhones, aiming to reduce latency, enhance privacy, and support offline use. PrismML has launched Bonsai 27B, using 1-bit and ternary weight techniques to compress a 27-billion-parameter model into 3.9GB–5.9GB, enabling on-device inference.

OpenAI has introduced GPT-Red, an internal automated red-teaming tool designed to大规模 identify prompt injection vulnerabilities and strengthen defenses before deployment. After adversarial training with GPT-Red, GPT-5.6 Sol shows approximately six times greater robustness against prompt injection attacks compared to the best production model from four months prior. OpenAI believes AI agents and tools like GPT-Red can form a self-improving 「safety flywheel」, making future models more robust, aligned, and trustworthy. The release was disclosed by Greg Brockman and others. This comes amid rising AI agent security incidents, including past cases where prompt injection attacks induced agents to leak private repository files.

OpenAI and Work Louder have jointly launched the Codex Micro hardware keyboard, priced at $230. The device features RGB agent status indicators, a joystick for workflow shortcuts, and a knob to adjust reasoning intensity. Users can press and hold the Micro key to verbally dictate full coding task requests without touching their laptop keyboard; the analog joystick can be mapped to common skills or Plan mode, while the knob toggles between fast tasks and deep reasoning. Fixed tasks become part of a hardware-based multitasking interface, with dedicated keys to switch tasks, check execution status, and approve permission requests. All controls can be customized to map to common Codex operations such as invoking skills or creating pull requests.

Perplexity has developed its own sandbox lifecycle management platform called SPACE, replacing third-party providers. It now handles 100% of Perplexity Computer's production traffic, achieving a fivefold improvement in tail latency. Designed for long-running agent sessions involving sensitive files, SPACE never directly stores credentials in its runtime—an essential security feature for extended agent trajectories. The platform creates isolated code execution environments for each task, forming a foundational infrastructure for reliable agent operations.

xAI announced the open-sourcing of Grok Build and shared its code repository, while resetting usage quotas for all users. Industry observer Elvis praised the move and advocated that all AI harnesses should be open source, recommending developers build and maintain their own harnesses. This action continues the recent trend of open-sourcing AI infrastructure and development tools.

In July 2026, Apple and Spotify launched new AI voice features. From July 13 to 14, Apple opened the iOS 27 beta to the public, featuring a redesigned Siri AI with screen awareness capable of understanding current screen content and leveraging personal data such as emails and photos to perform actions within supported apps. Initially available only in English and excluding the EU region, this update marks a significant advancement in contextual AI assistance. Shortly after, on July 14, Spotify launched the test version of 「Talk to Spotify」, allowing users to interact via text or voice to receive music recommendations based on personal listening history, get song explanations, or control playback. Currently limited to Premium users in the U.S., Ireland, and Sweden with English interfaces, this enhancement signifies deeper integration of conversational AI into mainstream applications to boost user engagement.

Google has introduced four enhancements to Gemini Spark: support for Google Docs editing, deeper Workspace integration, improved performance, and parallel source processing. As a 24/7 background-running AI agent, Gemini Spark is now rolling out to Google AI Ultra subscribers in additional countries and languages.

Reports indicate DeepSeek plans to go public by the end of 2026 or in 2027, having previously explored raising $1.5 billion at a $71 billion valuation. Concurrently, a trend toward financialization of AI compute is emerging: Kalshi has launched prediction market-based instruments to construct forward price curves for AI compute, advancing the commoditization of computational resources. Elon Musk acquired APR Energy to support the energy demands of large-scale Grok models.

Google DeepMind CEO Demis Hassabis proposed establishing a public-private frontier AI standards organization to test advanced AI models before deployment in the U.S., granting exemptions for non-frontier models, implementing quarterly benchmark assessments, and installing a 「kill-switch」 for severe risks. He called on the U.S. to lead in forming a global AI standards body. This proposal emerges amid growing discussions on AI safety regulation, with New York becoming the first U.S. state to ban the construction of new large-scale AI data centers.