AI News You Missed #110: OpenAI Makes Work the Product, Grok Comes Back Through Cursor, Meta Closes the Gates

1. OpenAI makes work the product

On July 9 OpenAI shipped GPT-5.6 in three tiers, Sol, Terra, and Luna, priced at $5/$30, $2.50/$15, and $1/$6 per million tokens, after completing a US government review, the same gate that held Fable 5 in June. The bigger move is the rebrand around it: ChatGPT Work turns the chatbot into an agent that gathers context from your apps and files and delivers finished documents, spreadsheets, presentations, and websites, with Codex folded into a single desktop app. GPT-5.6 also became the preferred model in Microsoft 365 Copilot. Meanwhile Anthropic's Fable 5, restored globally on July 1, moved to usage credits once subscribers pass their plan's included amount, and Google delayed Gemini 3.5 Pro past its June target with no new date.

Both US flagship models now ship through a government review. Regulatory clearance has quietly become a step in the release cycle, like an App Store review for frontier AI.
The product is no longer chat. It is finished work, priced by tier, distributed inside the tools your teams already use.

2. Grok comes back through Cursor

SpaceXAI released Grok 4.5 on July 8, its first model since going public and agreeing to acquire Cursor for $60B. It was trained jointly with Cursor on trillions of tokens of real developer sessions, priced at $2/$6 per million tokens, roughly 60% below Opus 4.8 on input, and pitched by Musk as "an Opus-class model, but faster, more token-efficient and lower cost." Benchmarks are self-reported and mixed (independent rankings put it fourth, and Cursor disclosed that a snapshot of its own codebase leaked into training), but the distribution is the point: it ships as a default option inside a tool millions of developers already live in. EU availability is expected mid-July.

Cursor is the Trojan horse. Nobody has to choose Grok; it is simply there, cheap, where the work happens.
At near price parity with GLM and other low-cost Chinese models, the main reason to adopt those models (price) largely evaporates, without the compliance questions.

3. Meta closes the gates

Meta shipped Muse Spark 1.1 on July 9 alongside a public preview of the Meta Model API at $1.25/$4.25 per million tokens, the first model Meta has ever charged for, and a closed-weight one. Muse Image, its first Superintelligence Labs image model, landed two days earlier. Zuckerberg posted on X for the first time in three years to sell it as "a strong agentic and coding model at a very low price," and Meta says the Muse family will replace Llama across WhatsApp, Instagram, Facebook, and its glasses. The company that spent three years arguing open weights would win just put a meter on its frontier.

The Llama era is effectively over. Enterprises that built on Meta's open models should re-check their assumptions about what stays free and open.
Meta's real asset is not the benchmark score. It is distribution across three billion users, now connected to a paid API.

4. Beijing weighs fencing in models

The hard news is Alibaba: it banned Claude Code for all employees effective July 10, after researchers found China-detection code inside the tool and weeks after Anthropic accused Alibaba-linked operators of a 25,000 account distillation campaign. The larger story is still in the meeting room: Reuters reported on July 7 that Chinese authorities have met with Alibaba, ByteDance, and Z.ai over the past month about potentially restricting overseas access to China's most advanced models, including unreleased ones. Nothing is decided, no mechanism exists, and sources say any curbs might apply only to future models. But the direction of travel mirrors #109's lead in reverse.

Separate fact from signal: the Alibaba ban happened; the export curbs are talks. Similar talks preceded action in Washington, so continued access to low-cost Chinese models is an assumption worth stress-testing, not a given.
Both moves came outside published rules, through an internal directive and closed-door meetings. Even before any policy lands, predictability is already the casualty.

5. Memory out-earns AI labs

Samsung's preliminary Q2 showed 89.4 trillion won (about $58.5B) in operating profit, up 1,810%, passing Nvidia's $53.54B to become the world's most profitable company for the quarter. The stock still fell about 7% as investors asked how long memory pricing can hold. Its chip division expects 2026 to out-earn its entire 40-year semiconductor history. Three days later SK Hynix listed on Nasdaq, raising $26.5B at $149 per ADR in the largest US share sale ever by a foreign company, seven times oversubscribed, opening 14% up.

While token prices collapsed this week, the profit pool moved down the stack. The scarce asset is no longer the model; it is the memory underneath it.
A record quarter met a falling stock. Markets are already pricing the question every capex committee should ask: is this a supercycle or a peak?

Report

Robert Maciejko: Reflections on corporate governance. Reporting from Nashville alongside roughly 900 governance leaders at the Society for Corporate Governance conference. Board caution on AI is professionals paid to imagine what could go wrong doing exactly that, yet directors expected to govern AI often have no sanctioned way to touch it. The question has shifted from whether boards engage to how fast and with what safeguards.

Bring the week's signal to your board.

If your board is wrestling with agentic work in the tools, model distribution through developer platforms, the return of closed weights, or memory and compute concentration as the profit pool shifts, start with a conversation.

Let's Talk Request a Session Invite Robert to Speak

AI News You Missed #110