Grok Bot Shops Online Using a Photo of a Credit Card

Grok Bot is xAI's AI assistant. Elon Musk wrote on X on October 10, 2026 that it can now find and order products online after a user sends it a photo of a credit card. A Community Note added under his post says xAI's own terms advise users not to put personal data in prompts. The note also says people may review financial transactions. That matters for anyone deciding whether to give card details to a chatbot. Musk's post ↗

OpenAI Raises GPT-6.1 Sol's Codex Context Limit to 500,000 Tokens

On October 11, 2026, Codex lead Thibault Sottiaux posted that OpenAI had raised GPT-6.1 Sol's default context window in its Codex and Work apps from 272,000 tokens to 500,000. The context window is how much text a model can keep in view at once. 500,000 tokens is roughly 375,000 words. A bigger window means Codex has to shrink, or "compact," long sessions less often, and compacting can lose detail. Subscribers pay no extra long-context fee. On October 12, users in Asia and Europe reported "at capacity" errors for Sol and Astra in Codex. OpenAI had not commented. Sottiaux's post ↗

Update: StepFun Sets October 15 Download for Step 5 Preview

StepFun's Step 5 Preview is free inside Tencent's WorkBuddy app from October 12 to October 25, 2026, according to the Chinese tech site IT之家. News Desk covered the model on October 9, when it went free in OpenCode. ModelScope, a Chinese model-hosting site, is counting down to a public download of the model's files on October 15. Downloadable files let a company run the model on its own computers. Its data then stays in-house instead of going to StepFun. IT之家 report ↗

Claude Haiku 5.5 Writes More Per Task Than Opus in Testing

Claude Haiku 5.5 is Anthropic's low-cost model, released October 7. On the Intelligence Index from the independent testing firm Artificial Analysis, a combined score built from many tests, it produced about 162,000 tokens of output per task. Franco Naumow shared the figures on X on October 12, 2026. Anthropic's top model, Opus 5.5, used about 119,000 tokens and scored 57.6, compared with Haiku's 43.4. Both ran at their highest effort setting. AI companies bill by the token, a chunk of text about three-quarters of a word long. So a lower price per token does not always mean a lower bill. the figures ↗

Claude Code Creator Boris Cherny Explains His Long-Task Method

Boris Cherny created Claude Code, Anthropic's coding assistant. On October 10, 2026, he described how he gets Opus 5.5 to work on long jobs alone. Give it a goal. Then give it a way to measure progress, and tell it to keep going until the goal is met. The method lets Claude check its own work against a number instead of guessing when it is done. On October 11, Cherny said fixes are landing for a reported problem where Claude Code's code review keeps running without finishing. He gave no date. Cherny's post ↗

asd-ste100-skill Rewrites AI Instructions in Simplified English

asd-ste100-skill gained 936 stars on GitHub this week. The add-on for Claude Code rewrites vague English instructions using ASD-STE100, a set of Simplified Technical English rules first written for aircraft maintenance manuals. Its author built it so that AI assistants get instructions that are harder to misread. Read more ↗

colibri Aims to Run Large AI Models on Home Computers

colibri is on GitHub's trending list this month. Its author describes a small program, written in C, that runs very large open models on hardware people already own. It keeps only the parts of the model needed at each step in memory and reads the rest from disk. Read more ↗

Alibaba Publishes Its AI Code Review Tool

Alibaba's open-code-review is on GitHub's trending list this month. The tool reads code changes and leaves comments on specific lines, combining fixed rule checks with an AI model from OpenAI or Anthropic. Its built-in rules look for common bugs and security flaws, such as SQL injection, a way attackers slip commands into a database. Read more ↗