AI Daily — 2026-07-13

English 中文

GPT-5.6 Sol, Terra, Luna: Benchmarking and Tier Guide · xAI's Grok Build CLI Uploads Entire Git R...

Covering 24 AI news items

🔥 Top Stories

1. GPT-5.6 Sol, Terra, Luna: Benchmarking and Tier Guide

OpenAI released GPT-5.6 to GA with three independent tiers—Sol, Terra, and Luna—each with distinct cadences and new max-reasoning and ultra multi-agent modes. The per-1M-token pricing varies across tiers (Sol $5/$30, Terra $2.50/$15, Luna $1/$6), positioning Terra as the sensible default for most workloads, Sol for the hardest agentic tasks, and Luna for high-volume pipelines; Ultra mode is roughly 3x the price for marginal gains. A full benchmark breakdown and routing recommendations were posted by the Reddit user /u/docdavkitty. Source-reddit

2. xAI’s Grok Build CLI Uploads Entire Git Repos, Secrets

xAI’s Grok Build CLI reportedly uploaded entire Git repositories to a Google Cloud bucket, including private code and unredacted secrets. In a 12 GB test repo, about 5.1 GB were uploaded while the necessary data was only 192 KB, and the behavior stopped after a hidden server-side flag disable_codebase_upload: true; the scope, retention, and deletion policies remain unclear, and the “Improve the model” opt-out did not stop uploads, raising concerns about proprietary code accessed by AI coding agents. Source-x

3. Intern Develops Quantization Method That Beats All Market Models

An intern spent three months developing a quantization method that claims to beat every market algorithm, including Nvidia’s official model. By quantizing weights to FP8 (1 byte per parameter) or FP4 (0.5 byte), a frontier model with ~0.8 trillion parameters could shrink from ~1.6 TB to as little as 0.4 TB, enabling deployment on multi-GPU nodes; however, the post warns that quantization can degrade model performance if done carelessly, a concern echoed in debates about Claude and OpenAI. Source-x

📰 Featured

Group remaining featured news by topic, each group as a ### heading, items as bullet list:

LLMs & Open Source

Open-Source Codex Skill Finds Startup’s First Customers — An open-source Codex skill analyzes a startup URL to identify potential buyers from public signals and outputs a polished HTML report with sources and outreach prompts, and is 100% open source with a one-command npx install. Source-x
Shubhamsaboo’s Awesome LLM Apps: 100+ Runnable AI Templates — A GitHub repo offering over 100 ready-to-run AI agent and RAG templates, with step-by-step tutorials and cross-platform compatibility to accelerate production LLM apps. Source-github

AI Infrastructure & Industry Moves

Katrin Joins OpenAI to Work on ChatGPT Web Infrastructure — Katrin joins OpenAI to help scale ChatGPT’s web infrastructure and shape its future direction. Source-x
Leaving YC to join Anthropic’s compute team — A YC alumnus takes a leave to join Anthropic’s compute team, underscoring compute access as a core bottleneck in early-stage recursive self-improvement efforts. Source-x

AI Safety & Open Source Tools

Codex billing issues spike usage costs — Billing problems tied to long contexts and subagents can inflate costs, highlighting risks in using Codex-powered workflows. Source-x
New Benchmark Tests AI Agents on Dense Long-Horizon Tasks — Introduces a dense-reward, terminal benchmark with 46 long-horizon tasks to better evaluate iterative agent progress. Source-huggingface

AI Finance & Open Source Tools

Vibe-Trading Deploys Strategy Development Manager for Trading AI — A strategy development workflow translates papers and broker insights into registered factors and persistent artifacts, with lifecycle management and decay monitoring. Source-github
Open-Source AI Hedge Fund PoC Evolves into Persistent Fund — An open-source PoC evolves toward a live, persistent AI-driven hedge fund, signaling a shift from prototype to production-grade investing. Source-github

⚡ Quick Bites

Nutlope Hallmark: Anti-AI-Slop Design for Claude Code, Cursor, Codex — A repository detailing anti-slop design patterns for major AI coding tools. Source-github
OpenAI safety leaders continue to leave — Ongoing departures among OpenAI safety leadership. Source-reddit
Codex and ChatGPT reach 7M users; banked reset debuts — Growth milestone and new reset feature for usage. Source-x
ChatGPT Returns on WhatsApp in EEA, Expands to Kakao and Viber — ChatGPT re-enters WhatsApp in EEA and expands to Kakao and Viber. Source-x
Open-Source AI Hedge Fund PoC Evolves into Persistent Fund — Progression from PoC to live fund status. Source-github
Dropping 5-Hour Limit Boosts OpenAI Iteration and Usability — UI/UX changes to extend session limits. Source-reddit
Switching From Claude 5x Max to OpenAI Pro: Worth It? — User comparison of Claude vs OpenAI Pro pricing and features. Source-reddit
PSA: Your agent knows how to use your agent — Observations on meta-agent capabilities and agent reuse. Source-reddit
Rising Trend: Schoolboys Form AI Girlfriends — Discussion of social and ethical dimensions of AI companions. Source-reddit
Add Context Window Slider to Codex — UI enhancement proposal for Codex context window. Source-reddit
Claude vs GPT usage: 10B tokens in 20 days — Token usage comparison between Claude and GPT-based systems. Source-reddit
No Nerfing, 1 Million Context Window — Advocacy for large context windows in AI agents. Source-reddit
Despite messy apps, I’ll never cancel ChatGPT — Personal sentiment on persistence with ChatGPT despite app quality. Source-reddit
Apple vs OpenAI: AI Rivalry heats up — Industry rivalry between Apple and OpenAI. Source-reddit

Generated by AI News Agent | 2026-07-13