News8/23/2026
Charlie Marsh stated that he redesigned the startup lifecycle in the latest version of the Codex CLI, enabling tools to respond instantly and reportedly improving startup speed by approximately 25 times.
Charlie Marsh stated that he redesigned the startup lifecycle in the latest version of the Codex CLI, enabling tools to respond instantly and reportedly improving startup speed by approximately 25 times.
Ollama announced that the Kimi K3 is now available to all Pro and Max subscribers.
OpenAI Developers announced that developers can now view usage and expenditure by API key in the console.
Thinking Machines will make Inkling available for free for several weeks while it's available on OpenRouter, but only for use initiated via Agent harness.
OpenAI announced that it will reduce the price of GPT-5.6 Sol's API and credits by more than 20% over the next three months, while the usage included in subscription packages will remain unchanged.
Cursor staff confirmed that the promotion of including a Cursor Ultra with the SuperGrok Heavy has ended; users who have already successfully claimed the Ultra can retain their benefits.
Vercel Developers has launched is-agentic.com to evaluate how easily AI agents can read and operate public websites, including over 100 checks, usage visualizations, and fix tips.
Google Antigravity introduces Remote Control, allowing users to connect to Antigravity sessions running on other machines from their browsers, iOS, or Android devices. The rollout will begin with Ultra users.
Fish Audio has announced that it will extend its free API policy for S2.1 Pro to developers until November.
NVIDIA claims its AVO Agent system achieved 100.00 RHAEs in the ARC-AGI-3 open set, completing all 183 levels across 25 open environments.
Google DeepMind has disclosed a phased AI research program with Fenris Creations around the world of the game EVE, focusing on long-term planning, memory, continuous learning, and multi-agent behavior.
DeepSeek Harness v0.1.1-rc.1 adds native image input for DeepSeek-V4-Flash-Vision-Exp to the official DeepSeek adapter.
Tibo, product lead at OpenAI, said that after the team checked several reports of Codex usage anomalies, they found that many affected users were using the sub2api service, which converts subscriptions into APIs.
Doubao PC version's work task mode has launched the skills, connectors, and work partner features, currently offering more than 200 skills and connectors, and supporting user-defined skills.
OpenAI's product manager announced that GPT-Image-2 can now generate transparent background images in ChatGPT and API.
Anthropic researchers proposed analyzing the interference weights in neural networks based on their effectiveness and helpfulness, and observed that most high-impact weights are helpful to the model's behavior.
Anthropic released the "AI-Native SDLC Playbook," which describes software development as a cyclical process of planning, designing, building, testing, deploying, and maintaining.
NovelAI Diffusion V5 Curated and V5 Full are now available for use.
Anthropic has integrated Claude Mythos 5 with Claude Security, making it available for public testing to enterprise users. Scans are billed on a standard token basis, without the need to purchase additional components.
DeepSeek has launched its experimental visual understanding model, DeepSeek-V4-Flash-Vision-Exp, on its API platform, supporting JPEG, PNG, GIF, and WebP formats.
SenseTime has officially open-sourced its lightweight, native unified multimodal large model, SenseNova U1.5 Lite. The model focuses on realistic visual creation workflows, with key enhancements including native support for 3-4k context lengths for ultra-long instructions, improved visual quality, text and layout, native 4K high-resolution output, native image editing, and fine-grained visual control.
OpenAI's open-source AI Agent underlying execution framework, Codex Harness, handles the entire execution loop, including task understanding, memory retention, tool invocation, and human approval. It supports embedding into business systems via CLI, SDK, and app-server. Key features include sandboxed secure execution, native human-in-the-Loop, streaming event transmission, and a model-agnostic architecture.
Alibaba Tongyi has launched the Qwen-UI-Agent GUI intelligent agent foundation model, covering mobile phones, computers, web pages, and deep search scenarios. The model is trained on over 100 real devices and 150+ applications, outperforming flagship models such as GPT, Claude, and Gemini on multiple benchmarks including MobileWorld, OSWorld, and WebArena. The model has strict security boundaries, rejecting illegal requests and proactively verifying users before sensitive operations such as payments and deletions.
At the 2026 World Robot Conference, Lumos Robotics launched its embodied intelligence evolution engine, Lumos NexCore, positioned as an industrial "skill infrastructure." The platform integrates data assets, model training, evaluation and verification, skill encapsulation, and equipment operation, enabling robot capabilities to be invoked, trained, and iterated like cloud services.