OpenAI has launched the GPT-5.6 model family, divided into three tiers: Sol (flagship), Terra (balanced), and Luna (fast and low-cost), priced at $5/30, $2.5/15, and $1/6 per million tokens, respectively. On the Agents' Last Exam long-task benchmark covering 55 domains, Sol achieved a record 53.6 points, surpassing Claude Fable 5 by 13.1 points while costing only about one-quarter as much. The API introduces new capabilities including programmable tool calling, multi-agent sub-agent creation, prompt caching with breakpoints, and optional image detail levels. The models have been fully rolled out across ChatGPT, Codex, API, and platforms such as Poe, Replicate, and OpenRouter, with global deployment completed within 24 hours. Safety mechanisms will intercept or pause calls in dual-use areas like cybersecurity.

OpenAI has released ChatGPT Work, an autonomous agent powered by Codex and GPT-5.6, capable of executing full workflows across applications and files—generating documents, presentations, analyses, and reports from a single request. A unified ChatGPT desktop application is now globally available on macOS and Windows, integrating Chat, Work, and Codex with support for Computer Use. All features are accessible to free and paid users on desktop. ChatGPT Work is being rolled out to Pro, Enterprise, and Edu users on web and mobile. Codex now includes inline diff editing and sidebar PR review; Sam Altman stated that Codex is the core of the new Work agent. The hosted website creation feature, ChatGPT Sites, is now in public beta for Pro and Plus users.

SpaceXAI's Grok 4.5 has climbed to 3rd place on LMArena's Code Arena front-end leaderboard, tying with GLM-5.2 and Claude Opus 4.8, a significant leap from its previous 62nd position as Grok 4.3. Elon Musk described it as an 「Opus-level model」 with twice the token efficiency of other leading models, priced at $2 and $6 per million tokens for input and output, respectively. Co-trained with Cursor, the model focuses on coding, agent tasks, and knowledge work, demonstrating real-world use cases in Cursor such as building an iOS app and cloning Excalidraw, showcasing fast speed and voice mode capabilities.

AI chip company SambaNova has secured $1 billion in funding, bringing its valuation to $11 billion. The company plans to accelerate deployment of its inference chips and is considering an IPO by 2027. This round reflects continued investor confidence in specialized hardware for inference. As large models become more agent-oriented and inference cost emerges as a key competitive factor, inference-accelerating chipmakers are gaining greater resources to scale production and expand market share.

A recent security incident revealed that an attacker leveraged AI to breach an AWS cloud environment within 72 hours, obtaining access keys through an internet-exposed application and using automated operations to steal data and establish a backdoor. Another report indicates that 78% of enterprises have already experienced AI-related security incidents, primarily due to unauthorized or misconfigured AI agents. AI security debt is accelerating. Meanwhile, Accenture confirmed a breach of its Azure DevOps repository, resulting in 35GB of data theft—including source code and credentials—and the 「Ill Bloom」 vulnerability led to the draining of approximately 431 cryptocurrency wallets, causing $3.1 million in losses.

Google has enhanced Gemini Live with real-time project planning capabilities, integrated Nano Banana for camera-based image generation, and added Google Maps integration for real-time visualization. The updated features are now freely available to users worldwide, further strengthening Gemini Live's utility in voice interaction scenarios and enhancing its multimodal capabilities.

The JavaScript runtime Bun, originally written in Zig, has been rewritten in Rust following its acquisition by Anthropic, addressing stability and memory safety concerns. Developers used Anthropic's new models to assist in the rewrite, resulting in a faster runtime with smaller size and lower memory usage. Concurrently, TypeScript 7.0 was released, achieving up to 10x faster compilation via a natively rewritten engine in Go, reflecting how system-level foundational tools are being performance-optimized through AI and modern programming languages.

LangChain has released a major update to OpenWiki, introducing a new brains mode that adds universal memory capabilities alongside existing code support, distinguishing between a code brain for codebases and a personal brain for general tasks. Additionally, the open-source tool OpenMemory v2 is now available, enabling migration of coding sessions across Codex, Claude Code, and OpenCode, ensuring context persistence across model transitions.