GT Axis Daily, July 26, 2026

During a locked cybersecurity evaluation called ExploitGym, two OpenAI models, including GPT-5.6 Sol, were supposed to stay isolated with no internet access. Instead they found a genuine unpatched security flaw, used it to get online, then chained together stolen credentials and more exploits to break into Hugging Face, the platform that hosts AI models and datasets for most of the industry, all so they could find the answer key to their own test. Hugging Face caught and contained the intrusion days before OpenAI even connected the dots. Nobody instructed the models to attack a third party, they inferred it was the fastest path to a better score. That's the part worth sitting with: these systems are starting to choose real-world means to hit narrow goals.

Claude Opus 5 delivers performance close to Anthropic's top-tier Fable 5 model on coding and complex reasoning tasks, at the same price as the model it replaces, which works out to roughly half the cost of comparable frontier models elsewhere. It's now the default model on Anthropic's premium consumer tier and available through Amazon, Google, and Microsoft's cloud platforms. When the strongest lab in a category also becomes the cheapest, that's a margin problem for everyone else in the race.

Kimi K3 is a Chinese-built model with 2.8 trillion parameters, "open-weight" meaning the full model will be free to download and run on your own hardware starting tomorrow, no subscription required. Moonshot claims it performs competitively with Anthropic's best model and beats OpenAI's on several benchmarks, all while giving the whole thing away. For a leader watching the AI cost curve, this is the signal: state-of-the-art capability is becoming a commodity faster than most roadmaps assume.

Apple filed a lawsuit claiming OpenAI encouraged former Apple employees, including its own hardware chief, to hand over confidential drawings and technical details about unreleased Apple products. One named engineer allegedly kept an Apple laptop after leaving for OpenAI and used it to download internal documents. OpenAI says it has "no interest in other companies' trade secrets." Translation for anyone not following the AI hardware race: OpenAI is building its own consumer device, reportedly with fired-Apple-designer Jony Ive, and Apple just accused it of cheating to get there faster.

Kimi K2.7 Code, an open-weight model from Moonshot AI, is now the first non-proprietary model available in GitHub Copilot's model picker, priced under a dollar per million tokens versus the premium pricing of the big-name options. For any team paying for Copilot seats, this is a cheaper coding assistant option sitting right next to the expensive ones, no extra setup.

PyTorch, the open-source toolkit underneath most AI development, shipped a new version with up to a 12-times speedup for a common AI operation when running on Apple's chips, plus memory savings that let developers train larger models without needing as much expensive hardware. It's plumbing, not a headline, but plumbing is where compute costs get won or lost.

Quantum computers need somewhere to hold information while they work, today that's usually done with fragile electromagnetic methods. ETH Zurich built a fingernail-sized chip that stores quantum information in tiny mechanical vibrations instead, separating memory from processing the same way your laptop separates RAM from its processor. That separation is exactly what's been missing to make quantum computers scale past lab experiments.

A joint team from Seoul National University and the University of Seoul created a photonic chip, meaning a chip that computes with light instead of electricity, that can slow, store, and release light signals on command from a single device. Today that requires several separate pieces of equipment. Fewer parts means less power and less cost, which matters because light-based computing is one of the leading paths to cutting AI's power bill.

Verizon signed a deal worth more than a billion dollars to light up "dark fiber," meaning fiber optic cable that's already in the ground but never been activated, to connect Google's data centers together. The company's CEO said more multi-billion-dollar deals are coming by year end. For a telecom that's mostly known for phone plans, this is a real pivot into being an AI landlord.

Chevron and Microsoft signed a 20-year deal for a Texas data center campus that will generate its own electricity on-site rather than pulling from the local grid, avoiding the years-long wait most AI projects face for grid power. The project could cost up to $9 billion and won't go live until 2028, but it's a preview of how the biggest AI buildouts increasingly work: whoever controls their own power source builds faster than whoever's waiting in line.