Introduction

Here are six notable stories selected from overseas AI news collected on July 29, 2026. Today's lineup is a rich one: OpenAI resuming usage limits, structural change in the software industry, a new model distillation method, progress in edge AI, LLMs applied to security tooling, and questions about benchmark reliability.

---

1. ChatGPT Work and Codex 5-Hour Limits Resume "Tomorrow" — Addressing GPT-5.6 Sol's Token Consumption Issue

On July 29, OpenAI's Thibault Sottier announced that the five-hour usage limit windows for ChatGPT Work, the AI service for desktop PCs, and Codex, the AI coding tool, would resume the following day.

Following the release of GPT-5.6 Sol, token consumption far exceeded expectations and the limits were temporarily relaxed. The resumption suggests that consumption optimization has progressed to a certain degree. GPT-5.6 Sol's strong reasoning capabilities may be changing how users work, so continued monitoring will be necessary.

On the practical side, users who run into the limits during coding work may want to start thinking about streamlining their prompts and splitting up their tasks.

2. AI Is Changing "What Customers Buy" — An Existential Crisis for Software Companies

According to Bizjournals, software companies are confronting business model changes driven by advances in AI. Customers increasingly want "outcomes" rather than "tools," and the traditional license-sales model is starting to crumble.

This is not merely a matter of declining revenue — it is forcing software companies to fundamentally rethink what kind of company they are and what they offer. The backdrop is that as AI agents become capable of completing tasks autonomously, users have less need to piece together individual SaaS tools.

Industry groups and analysts are divided over how quickly this structural change will play out, but at the very least, only a limited number of companies will be able to afford to postpone a rethink of their product strategies.

3. Relay-OPD: Fixing "Prefix Failures" in Reasoning Models with On-Policy Distillation

In a paper featured on Hugging Face Daily Papers, a research team including Zhejiang University proposed a new knowledge distillation method called Relay-OPD.

In conventional on-policy distillation (OPD), when a small model (the student) heads in the wrong direction early in its reasoning, the entire subsequent rollout ends up built on mistaken premises. Relay-OPD detects online the moment the teacher's and student's outputs diverge, has the teacher temporarily take over at that point, and then hands the baton back to the student.

Training 0.6B/1.7B student models with Qwen3-4B as the teacher, the method achieved +5.73% over standard OPD and +1.49% over FastOPD, the previous best, across eight mathematical reasoning benchmarks, while cutting training trajectory length by more than 50%. It is also notable for its practicality: no additional verifiers or reward models are required.

4. BetterGPT-150M: A 150M-Parameter Small Model Surpasses GPT-2 Small

A language model called BetterGPT-150M, with roughly 152 million parameters, has been released in the LocalLLaMA community on Reddit. It was pre-trained on 15B tokens using curated datasets including FineWeb-Edu, fine maths, cosmopedia, and starcode-python.

In benchmark evaluations, it outperforms GPT-2 Small while achieving results comparable to models trained on similar token budgets. It is a base model with no instruction tuning, but its low barrier to entry — it can run inference instantly on a CPU — should make it valuable for experimentation on edge devices and for educational use.

The weights and a demo are available on Hugging Face as open source, and feedback from the community is being solicited.

5. CodeCrucible: Block Publishes a Blueprint for LLM-Powered Static Security Analysis

The engineering team at Block (formerly Square) has published the design philosophy behind its LLM-powered static application security testing (SAST) as "CodeCrucible."

Traditional SAST tools suffer from many false positives, exposing the limits of rule-based approaches. Leveraging LLMs makes it possible to judge vulnerabilities with contextual understanding, which is expected to reduce false positives and improve detection accuracy. Block has systematized lessons from production use and shared them as design patterns that other companies can reference.

LLM applications in security are entering the practical stage, and with benchmarks such as ExploitGym now emerging, the ecosystem as a whole is maturing.

6. Benchmark Contamination in Multimodal Fact-Checking — 17–29% of "New Claims" Are Already Known

A paper accepted at ACM MM 2026 has flagged data contamination issues in multimodal automated fact-checking (MAFC) benchmarks.

The research team developed a "Contamination Detection Pipeline" and quantified the reality that LLMs/VLMs earn artificially high benchmark scores through memorization of their pre-training data. Surprisingly, even in "dynamic" benchmarks (which include new claims from after the cutoff), 17–29% of the claims could be verified with prior knowledge.

The paper shows that data contamination inflates Macro-F1 by up to 11.34 points and even distorts rankings between models. This is an important finding that calls into question the very assumption that a high leaderboard score equals real capability.

---