Don't use Claude Fable 5.2 at Max effort!!! Artificial Analysis benchmark results: High: 62 | $9.08/task | 17.7 min XHigh: 62 | $14.83/task | 26.3 min Max: 65 | $26.54/task | 46.4 min Going from High to Max gets you just 3 more points, but costs nearly 3x as much and takes 2.6x as long. For everyday coding, I'd start with High and save Max for the really hard tasks. The highest setting isn't always worth it.

Heads up if you've updated the Codex app. Check Settings > Data Controls > Location. If enabled, ChatGPT can access your precise location, potentially down to your exact address. Location data may also be shared with third-party providers for local results. It's supposed to be off by default, but worth double-checking.

A few months ago, Goldman Sachs was encouraging non-engineers to build tools with AI coding agents. Now it's pulling back. According to reports, most traders and investment bankers are no longer allowed to use tools like Claude Code. Access is largely limited to engineers and certain technical roles. And with the recent concerns around leaked API keys, exposed credentials, and agents accidentally deleting files, I can see why. When an agent has access to internal code, customer data, and company systems, a single mistake can do real damage. AI is making it easier for anyone to write code. But access controls, secret management, and code review matter more than ever.

AI coding agents are already pretty good at writing frontend code. Getting them to build a consistent UI across an entire project is still tricky. Meta's open-source design system, Astryx, takes an interesting approach to this. It lets agents like Claude Code and Codex access component docs, templates, and design guidelines through its CLI. Instead of making UI decisions from scratch every time, agents can build around an existing design system.

Coding agents are writing code faster than ever, but Git infrastructure is becoming a bottleneck. GitHub just introduced Project Rubicon, a new Git storage architecture built to handle this growing load. In internal tests, it supported 5,000 agents working on the same repository and handled 7,200 pushes per second. It separates compute from storage, so GitHub can scale without changing how developers use Git. OpenAI and Anthropic have also built their own systems, Pando and Scion, to tackle similar problems. AI is changing how fast we write code. Now the infrastructure needs to catch up.

An interesting uncensored model just showed up on Hugging Face: DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED. The developers first fine-tuned Qwen3.5-9B on around 250K math and science examples generated by DeepSeek-V4-Pro, then modified it to make it less likely to refuse requests. The results are interesting. In the creator's 100-prompt refusal test, the original model refused all 100. The modified version refused just 3. It's also a relatively small 9B model, with a Q4 quantized version around 5.6GB, making it possible to run locally on a regular computer. What I find interesting is that you can change how a model behaves without retraining the whole thing.

Anthropic's expected IPO valuation just dropped again. According to Bloomberg's November 6 report, potential investors are now discussing a valuation of $1.2T to $1.3T, with some going as low as $1.1T. Just a month ago, figures as high as $2T were being floated. The market is starting to take a harder look at AI valuations. Better models and growing revenue matter, but investors increasingly want to know how much profit these companies can actually make after spending billions on compute.