GPT‑5.6 has just been released with three new versions: Sol – the flagship, Terra – balanced for daily work, and Luna – the most cost‑effective.
Naturally, everyone can guess that their benchmark scores are sky‑high; this time they published a lengthy article to introduce the models and showcase what they can actually do.
GPT-5.6: Frontier intelligence that scales with your ambition
Cloudflare has introduced many features; Workers now have a built‑in cache that can be enabled with just a few lines in the Wrangler configuration and by adding a header.
You don’t need any extra libraries or tools—just a couple of code lines and the cache works immediately. Clearing the cache is also straightforward; you only need to call the purge function.
I’m not sure if anyone’s heard of this chrome-devtools-mcp yet, but it’s really impressive. Just hook MCP up to the agents and they browse the web like a god—anything I can do or see, they can do too.
Instead of copying those red messages from the console, I now just feed them straight into the reader 🤣
A paper describes an approach to the image‑search problem (Images for RAG). Typically, you just feed an image into an embedding model → obtain a vector → store it → query the vector → get the result.
Kapa, however, adds a simple vision model that generates a textual description of the image and stores that description in a database. Since the data is now plain text, there are many ways to query it, and it still produces very good results.
GitHub Copilot has just officially added an open‑source large language model to its service. The model is Kimi 2.7, from the Chinese AI company Moonshot. Compared with big names like Claude or ChatGPT, Kimi offers a very affordable price—much lower, especially now that Copilot has changed its usage‑based pricing for all models, which makes users feel more constrained than before.
I’ve spent some time testing Kimi 2.7 and it really is good; it can handle most daily tasks and simple programming. I described how I used Kimi to rebuild this blog in my latest post. How about you? Have you tried this model? What do you think?
Củng ko có gì lạ khi mấy mô hình từ TQ quá rẻ
Tại sao lại kp là deepseek trong khi giá của nó rẻ hơn nhiều mà tốt nữa
cline.bot was first known as a VSCode extension that provided code‑completion suggestions, gradually expanding into a chat feature and eventually into agents that run entirely within the code editor. Recently, Cline officially entered the provider race by launching ClinePass — a monthly subscription service that gives access to many open‑source models, either integrated into the CLI or accessed via an API.
At first glance it looks a lot like OpenCode Go’s model, which isn’t surprising because OpenCode just announced that its Go package achieved very high sales in two months. The difference is that OpenCode publicly states usage limits, while ClinePass has not disclosed specific limits; it only says the limits are 2–5 times higher than the regular pricing for each model type.
The price is $9.99 per month, $4.99 for the first month. There is currently a very attractive promotion: you can sign up for just $1.99 through the CLI, which is worth trying. To get the discount, install Cline with npm i -g cline, start the CLI by running cline, register an account, then subscribe to ClinePass to receive the coupon code. Note that this offer is only available when you create the account via the cline CLI.
An article from early 2026 by OpenAI is surprisingly valuable for scaling their PostgreSQL servers. The techniques it presents look simple, yet when combined they deliver a substantial boost.
I heard rumors that Claude Fable 5 is set to come back soon after being tweaked to nerf its power before it’s officially released. I’m excited, but the only thing I’m really interested in is whether I can get a free trial with the $20 package 😅
Mở lại thật rồi mọi người ạ, đến 07/07 và được dùng 50% limits của tuần. Sau đó là... trả tiền nha.
Claude Sonnet 5 is officially released; its performance is almost on par with Opus 4.8, yet its price matches the earlier Sonnet 4.6. There’s no need to wait for more information—just give it a try, everyone
Sao thấy mấy mạng xã hội khác bạn ẩn mình vậy? Ngoài cái website này bạn còn chơi cái nào khác nữa không?
À trước đây mình có dùng Threads nhưng không được thuật toán ưu ái, kèm theo việc vận hành nội dung trên đó tốn thêm thời gian nên mình quyết định không tiếp tục nữa. Thay vào đó thì tự tạo công cụ, tự viết, tự vận hành trang này vì nó nhanh và đáp ứng được thời gian biểu của mình á.
Nói ngắn gọn thì xây dựng nội dung trên mạng xã hội tốn khá nhiều thời gian, chịu sự chi phối của thuật toán mà mình không thích. Không phải có ý bài trừ, mà ý là nếu có những người cùng tư tưởng, nhận thấy giá trị thì họ sẽ quay lại thường xuyên hơn và có xu hướng gắn kết, dễ kết bạn và trao đổi hơn.
I just received an email from DeepSeek informing me that they will raise their API prices. Specifically, in mid‑July they’ll launch the official DeepSeek V4, which includes many optimizations and performance improvements. As a result, the API price will double during peak hours—namely the two windows of 01:00 – 04:00 UTC and 06:00 – 10:00 UTC.