GT Axis Daily, July 11, 2026

GPT-5.6 comes in three sizes: Sol (the flagship), Terra (the mid-tier), and Luna (the cheap and fast one). The rollout had been held back pending a US government review, and that review cleared on July 9. The headline claim is agent-style software work, meaning AI that plans a task, runs the commands, checks its own output, and iterates instead of just spitting out a paragraph. Sol runs $5 per million words in and $30 out, Luna runs $1 and $6. That spread is the point: you no longer buy "AI," you buy a tier.

Muse Spark 1.1 landed the same day as GPT-5.6. It handles a million words of context at once, meaning it can hold an entire codebase in working memory, and it can take screenshots of what it built to check its own work. Pricing is $1.25 in and $4.25 out per million words, roughly a quarter of what OpenAI's flagship charges. Meta is not trying to win the benchmark, it is trying to make the benchmark not worth paying for.

The European Commission issued a preliminary finding that Meta breached the Digital Services Act, the EU's rulebook for large online platforms. The specific targets are infinite scroll, autoplay, and push notifications, features the Commission says push users into compulsive use without Meta ever properly assessing the mental health risk. The proposed remedy is to turn those features off by default. The potential fine is up to 6% of global revenue, which on Meta's base is real money.

Between OpenAI's Luna at $1 per million words in, Meta's Muse Spark 1.1 at $1.25, and the rest of the field scrambling, the cost of putting an AI agent on a software task has dropped by roughly an order of magnitude in a year. Meanwhile GitHub shipped a desktop Copilot app and made its Copilot SDK generally available, meaning any company can now embed GitHub's agent engine directly into its own product. The tools are commoditizing. The workflow around them is where the value moves.

Cursor added Automations, meaning AI coding agents that trigger themselves off a GitHub pull request, a Slack emoji, or a schedule, then go do the work and open the pull request themselves. You describe the job in plain English and it wires up the trigger for you. This is the quiet shift nobody is naming: we went from "AI answers when I ask" to "AI acts when something happens."

The chip is codenamed Iris, designed in-house with help from Broadcom and manufactured by TSMC. It is Meta's fourth attempt at custom silicon and the first one that looks like it will run at scale. The strategic point: Meta plans to go from 7 gigawatts of computing capacity this year to 14 next year, and it does not want to write NVIDIA a check for every watt of it. Testing reportedly took six weeks with no major issues.

Rémi Cadène, who worked on Tesla's humanoid program, unveiled UMA, a French startup building a lightweight humanoid called Northstar. He says the company is already in conversation with 50 potential customers. Europe has been a spectator in the robot race. This is the first credible attempt to change that, and it is being run by someone who has seen the inside of the leading program.

SK Hynix, the world's second largest memory maker and the primary supplier of the high-bandwidth memory that sits next to every AI chip, listed American shares on July 10 under the ticker SKHY. The shares priced at $149 and closed at $168.01, a 13% first-day pop. The money funds a new fab campus in Korea and a $4 billion packaging plant in Indiana. This is the largest thing that happened in tech this week and almost nobody outside the industry noticed.

The round was led by Aramco Ventures at an $8.3 billion valuation, with NVIDIA participating. Together AI runs the infrastructure for companies that want to use open models, meaning AI they can download and run themselves rather than rent from OpenAI. Annual bookings crossed $1.15 billion, and Cursor is a customer. The company says it will grow its capacity roughly 50x over five years.