Here are the five most important stories from the overseas AI news we spotted on July 31, 2026. We've skipped topics that overlap with our earlier coverage (the breach found during Claude testing, Gemini Robotics 2, the Chrome vulnerability discovery, and the like) and focused on new developments.

GPT-5.6 "Optimizes Itself," Cutting Flagship-Level Costs to 1/13 in Four Months

OpenAI has cut prices on GPT-5.6 sharply. The small model "Luna" is down 80%, "Terra" 20%, and the reasoning model "Sol" gets a new "Sol Fast" mode that is up to 2.5x faster (with intelligence said to be unchanged). The notable part is the reason for the cuts: OpenAI reports that GPT-5.6 itself (Sol) analyzed production traffic and autonomously tuned load balancing and rewrote kernels in the company's Triton/Gluon languages, cutting end-to-end serving costs by 20%. Improvements to the draft models for speculative decoding also raised token generation efficiency by more than 15%.

Observers say the long-noted trend of "the same intelligence getting cheaper" is holding. Today's Luna now matches the flagship-level scores of the GPT-5.4 era four months ago, at roughly 1/13 the token prices (GPT-5.4 at $2.50/$15 versus Luna at $0.20/$1.20). That works out to a pace of about 2,000x per year, and there are now situations where OpenAI beats open models like DeepSeek, GLM, and MiniMax on cost per unit of fine-tune-free intelligence. Automatic review in ChatGPT and Codex is also reported to be moving from GPT-5.4 to Luna at about one-tenth the cost. One caveat: public benchmarks may to some extent be in the training data, and the harder a metric is to train on directly — Elo being one — the more gains of this kind should be discounted.

Thinking Machines Releases the Small Open Model "Inkling-Small"

Thinking Machines has released Inkling-Small, an open-weight multimodal MoE. It has 276B total parameters and 12B active, making it about a quarter the size of the original Inkling while aiming for comparable performance. Alongside text it natively handles audio and images, and it can inspect images in Python during inference. Major inference stacks such as vLLM, SGLang, and LM Studio are supported from day 0, and some configurations also support 1M context.

On benchmarks it scores 40 on the Artificial Analysis intelligence index, one point behind the flagship Inkling. It is strong on Humanity's Last Exam, GPQA Diamond, and SciCode, and it reportedly beats the larger Inkling on several coding tasks. It is weaker on some agentic tasks and factual knowledge, but some are calling it one of the practically important releases in this batch — an open model this efficient on just 12B of active compute.

56% of Cursor's Merged PRs Now Originate from Cloud Agents

AI coding agents are moving out of demo territory and into the center of real work. Cursor reports that as of last December, "one in ten merged PRs" came from cloud agents — a share that has now reached 56%. The company credits giving agents dedicated cloud PCs and the ability to improve their own environments. Cognition's Devin, for its part, demonstrated running macOS in the cloud via Slack and a web app to build and test native iOS apps with Xcode and the simulator, and it now supports stacked PRs on GitHub. With engineers saying things like "I still haven't set up a local PC," the impression is that the working environment itself is shifting toward the agent side of the cloud.

Big Tech's AI Investment Tops $1 Trillion in Total, While Some Warn of "Debt Dependence"

The Financial Times reports that the industry-wide AI spending rush has exceeded $1 trillion in total — a milestone framed not by any single company's numbers but by the aggregate. Individual large investments such as data center plans have been reported one after another, but this is the first coverage to mark the cumulative "$1 trillion" threshold.

Caution on the funding side is growing, however. The research outlet Grey Swan argues that "the AI trade now runs on borrowed money, and lenders are beginning to reprice that risk," drawing a relatively large response on Hacker News. The Financial Times also reports that Leopold Aschenbrenner, once the "golden child" of the AI trade, has stalled (recent reports had him moving to raise funding). Combined with the AI stock plunge and rising credit risk seen a few weeks ago, it reads as a stage where the market is once again probing the balance between growth expectations and sustainability.

Perplexity Launches "Projects" and Persistent Memory "Brain"; the Verdict on Memory Infrastructure Remains Split

Perplexity has evolved its "Spaces" workspace into "Projects." Equipped with shared files and a persistent memory called "Brain," it is positioned as something like an agent OS for multiple people. Whether memory infrastructure genuinely contributes to capability, though, is still contested. TurboPuffer reports that Mem0 moved more than 400 million agent memories from pgvector to its own platform, achieving p90 70ms and recall@10 of 97%. But some research shows more cautious results: with filesystem-based memory, retrieval cost can be halved without improving final answer quality, and storage quality can degrade depending on the managing agent. Persistent memory is maturing as a product feature, but its causal contribution to capability seems unsettled at this point.

---