Anthropic released a new API version, 2026-10-21. It's the first formal API release since 2023. This version removes sampling parameters like temperature. If you pass them in, you'll get a 400 error. Developers should take note.

OpenAI has discovered two critical vulnerabilities in the PHP model. One vulnerability, dubbed "PHP-Exploit," has a severity score of 9.2 and can be exploited without requiring authentication. This vulnerability affects versions 8.5, 8.4, 8.3, and 8.2, and users should upgrade

Anthropic's open preview of Claude Duet research. Daily summary email and calendar. Check connected applications every few hours. Send reminders only when you need to take action. Tasks are executed on a cloud-based virtual machine, so your phone and computer don't need opening

Tibo, on day 15 of the 28-day plan, we're delivering a reset. This 15-day period is either a reset or an improvement – and we're committed to delivering on that promise. Check your codex to see if the reset has been applied.

ChatGPT has launched a feature to lock chats. And this must be accessed via password. Of course, if you reset it. All the passwords you set up before will definitely be deleted in chats. Damn, this really is a privacy department. What do you think about this?

Deepseek open-sourced its own inference framework, Deepseek Infa, and according to him, the cache hit rate can reach 96%. Also supporting Nvidia, now it's truly fully open-source—not only has the code been open-sourced, but the weights have also been open-sourced,how do you think

Approximately 35 organizations, including Meta, NVIDIA, Alibaba, DeepSeek, and Moonshot, have made open-source models available on Hugging Face. The total computing power used, calculated based on current GPU rental prices, amounts to nearly $100 million.

Anthropic has Establish a public welfare fund, first investing 5% of shares. For every 50% increase in valuation thereafter, the board can add another 2%. Including the co-founders' commitment to donate 80% of the shares and about half of the employees to donate proportionally

Now, Linus is also using Claude Code to optimize the Linux kernel, and each release includes a mention of LLM.

X has implemented a new, unified subscription model. This combines the previous options of Block, Person, and X Premium, offering plans for $8, $30, $100, and $200 respectively.

Tibo stated that the 28-day commitment has already been met. Currently, the system is resetting usage limits for all global users. The first five days were either for improvements or resets, and this commitment has been fulfilled.

Tibo just made a bold commitment for Codex: For the next 28 days, the team will either ship one meaningful improvement every day, or issue a full usage reset. One or the other. Every day. That's accountability. Most teams promise roadmaps. This is a daily scoreboard.

OpenAI's new pricing structure: Pro 100: base Pro 200: now 10x Plus usage (down from 20x) Pro 500: $500/month, 25x Plus, exclusive Ultrafast Existing Pro 200 users get 62,500 credits worth $2,500 as compensation.

OpenAI launched dots: an always-on agent. It has its own cloud computer. Connects to 4000+ apps via plugins. Works for you around the clock. Included in Pro and Business Premium. First month: usage doesn't count against quota.

GPT-6.1 Sol is the real headline from DevDay. Intelligence near Astra. API price: $2/$10 per M tokens. Cached input: $0.10—95% cheaper than standard. DeepSWE: matches Astra. OSWorld 2.0: +7 points over previous gen. And an Ultrafast version coming with 8x speed.

OpenAI DevDay just happened. Here's what shipped: GPT-6.1 Sol: near-Astra intelligence at 1/5 the price dots: always-on agent with its own cloud computer, connects to 4000+ apps Pro 500 plan: $500/month, 25x Plus usage, exclusive Ultrafast access

Manus released 2.0. New Cascade agent framework: 23.2% fewer tokens, 28.2% faster, 32% lower cost. Desktop becomes Manus Studio—video editor + game dev environment. Automation now supports event triggers.

Anthropic released Claude Sonnet 5.5. 30%+ faster output. Same token price ($2/$10 per M). Up to 30% lower cost per task—it uses fewer tokens. Terminal-Bench 4.0: 10.3% → 70.6%. GDPval-AA v2.1: 1844, nearly matching Opus 5.5's 1846.

AMD is acquiring Fei-Fei Li's World Labs for ~$8.2B in stock. Deal expected to close by end of 2026. Fei-Fei Li becomes AMD EVP and Chief Scientist, reporting directly to Lisa Su.

TypeSafe AI reopened Jev registration. They expanded capacity. But removed the $5 free credit for new users—"a few abusers affected everyone's experience." Jev is a model that doesn't generate text—it outputs structured decisions ~100x faster than LLMs. The demand is real.