OpenAI Pulling Out of Cursor; Sony, Warner Sue Anthropic; Mystery Model Revealed

Today's AI Outlook: 🌤️

Cursor Gets Caught in the Musk-Altman Crossfire

OpenAI is pulling direct access to its models from Cursor by Nov. 12 following the coding platform’s reported $60B acquisition by SpaceX. Cursor customers who still want GPT models will need to supply their own OpenAI API keys and pay OpenAI directly rather than accessing the models through Cursor’s bundled subscription.

OpenAI cited contractual concerns tied to Elon Musk’s ownership, while Cursor CEO Michael Truell has pushed for a resolution. Musk responded that he “couldn’t care less,” according to The Rundown AI.

Why it matters

Cursor built its appeal around giving developers access to competing frontier models in one place. Ownership by a major AI rival introduces a new strategic problem because model providers must decide whether distributing through Cursor also strengthens a competitor’s platform.

The Deets

• OpenAI said its models will leave Cursor on Nov. 12.
• Cursor previously bundled GPT access directly into its subscription.
• Developers will still be able to use OpenAI models with their own API credentials.
• OpenAI said the change followed contractual cancellation rights triggered by Cursor’s sale.
• The Rundown AI reported that OpenAI represents about 5% of Cursor’s AI traffic, according to Cursor CEO Michael Truell.
• Anthropic has so far reaffirmed support for Cursor.

Key takeaway

Cursor’s value as a model-neutral coding hub now depends on how many AI companies remain comfortable distributing their technology through a platform owned by a rival.

🧩 Jargon Buster - API key: A private credential that lets software connect directly to another company’s model or service and bill usage to the account holder.

AI Profits Get a $160B Paper Boost

The AI boom is showing up in a surprising place: Big Tech’s income statements. Alphabet, Amazon, Nvidia and Microsoft collectively recorded more than $160B in “other income” last quarter, according to reporting cited by AI Secret, with rising valuations of AI-related investments helping push reported profits higher. Alphabet recorded $97.9B, while Amazon reached $53.4B.

These gains are tied to the increasing value of equity holdings rather than customers buying more products or cloud services, which means the headline profit numbers can look far stronger than the underlying cash picture.

Why it matters

Investors are watching companies spend enormous sums on data centers, chips and energy while some of those same AI investments produce large accounting gains. That makes cash flow, operating income and investment gains increasingly important to separate when judging the health of the AI boom.

The Deets

• The four companies recorded more than $160B in “other income” last quarter.
• Alphabet’s figure reached $97.9B, roughly double the prior quarter.
• Amazon reported $53.4B.
• The gains reflect rising values of investments, including stakes in private technology companies.
• AI Secret cited concerns from Goldman Sachs and Citi about whether current profit growth can continue if investment gains cool.

Key takeaway

AI investments are boosting corporate earnings before many of the underlying bets have produced corresponding operating cash. That makes the fine print below net income unusually important.

🧩 Jargon Buster - Paper gain: An increase in the value of an investment that boosts reported wealth or earnings even though the asset has not been sold for cash.

Anthropic’s Copyright Fight Gets Personal

Sony Music Publishing and Warner Chappell Music sued Anthropic, CEO Dario Amodei and co-founder Benjamin Mann in federal court on Aug. 28, accusing the company of using pirated material while assembling data for AI training.

According to the complaint described by both newsletters, plaintiffs allege that Mann personally downloaded millions of books from LibGen and that internal discussions acknowledged concerns about the source of the material.

Why it matters

Copyright litigation against AI companies has often focused on whether training itself qualifies as fair use. This case puts much more attention on how training material was obtained and whether individual executives can be held responsible for decisions surrounding those data pipelines.

The Deets

• Sony Music Publishing and Warner Chappell Music filed the lawsuit on Aug. 28.
• The complaint names Anthropic, Amodei and Mann as defendants.
• Plaintiffs allege Mann downloaded millions of pirated books from LibGen.
• The publishers are seeking statutory damages that can reach $150,000 per work.
• The case centers heavily on data acquisition practices rather than solely on what happened after the content entered a training dataset.

Key takeaway

AI copyright cases are moving deeper into the mechanics of data collection, placing executives and internal decision-making alongside model training itself under legal scrutiny.

🧩 Jargon Buster - Training data: The collection of text, images, code and other information used to teach an AI model patterns it later uses to generate outputs.

Anthropic Wins a Round Against the Pentagon

A federal judge ruled that the Pentagon acted unlawfully when it labeled Anthropic a “supply chain risk,” finding that the designation was retaliation for the company’s objections to military AI policies rather than a response to an established security threat, according to reports.

The designation had forced Anthropic models offline across government agencies and represented an unusually severe action against a U.S. AI company.

Why it matters

The ruling creates a legal obstacle for government agencies seeking to restrict AI vendors through national-security authorities when disagreements involve company policy, speech or acceptable-use restrictions.

The Deets

• The Department of War issued the designation in June.
• Government agencies were ordered to remove Anthropic models.
• Judge Rita Lin ruled that the action violated the First and Fifth Amendments.
• A separate appeals case means the designation has not completely disappeared.
• The Pentagon still plans to finish winding down Claude usage by Sept. 30.

Key takeaway

AI companies negotiating government contracts now have clearer legal grounds to challenge restrictions they believe are retaliatory rather than security-driven.

🧩 Jargon Buster - Supply chain risk: A government designation suggesting that a vendor, technology or supplier could create security risks within critical systems or operations.

🛠️ Tools & Products

Codex Takes On App Store Bureaucracy

OpenAI’s Codex can now help developers move an iOS app through much of the App Store submission process, according to a guide highlighted by The Rundown AI. The workflow connects development work in Xcode with App Store Connect, allowing Codex to assist with building, archiving, listing information, screenshots and submission preparation.

Why it matters

Coding agents are expanding beyond writing software into the administrative steps required to actually ship it. For independent developers, those tedious final-mile tasks can consume nearly as much attention as development itself.

The Deets

• Codex can build, archive and distribute an app from Xcode.
• It can assist with App Store descriptions and screenshots.
• Developers can use it while completing age-rating, contact and content-rights information.
• Codex can also help add Apple’s StoreKit framework for subscriptions or paid upgrades.
• Human review is still important before submitting information to Apple.

Key takeaway

Coding agents are becoming shipping agents, helping developers navigate the paperwork that sits between finished software and paying customers.

🧩 Jargon Buster - StoreKit: Apple’s framework for adding purchases, subscriptions and other paid features inside apps.

🧪 Research & Models

The Mystery Model Had a Price Tag

The anonymous Ox Alpha model that surged across OpenRouter and OpenCode has been revealed as GLM-5.3 Flash from Chinese AI lab Zhipu. During its anonymous test period, developers could use the model for free, helping it climb usage rankings before its creator was publicly known.

According to reporting, the model scored 57 on the AA benchmark, matching Claude Opus 4.8 while costing roughly 1/40th as much.

Meanwhile, Tencent released Hy4-preview, a 770B-parameter open-source Hunyuan model with 49B active parameters and a 1M-token context window.

Why it matters

Model competition is increasingly shaped by price, accessibility and deployment efficiency alongside raw benchmark scores. Strong performance at dramatically lower cost gives developers more freedom to route workloads away from premium frontier models.

The Deets

• Ox Alpha was revealed as Zhipu’s GLM-5.3 Flash.
• The anonymous model became heavily used on OpenRouter and OpenCode.
• AI Secret reported a score of 57 on the AA benchmark.
• Its reported pricing is about 1/40th the cost of Claude Opus 4.8.
• Tencent’s Hy4-preview has 770B total parameters, with 49B active at a time.
• Hy4-preview supports a 1M-token context window.

Key takeaway

The frontier is getting crowded with models that pair strong capability with much lower operating costs, giving developers more leverage when choosing what runs behind their applications.

🧩 Jargon Buster - Active parameters: The portion of a model’s total parameters actually used while processing a given request, which can reduce computing costs in mixture-of-experts models.

⚡ Quick Hits

• Texas puts the brakes on Flock: Flock Safety lost state funding for new AI license-plate cameras after officials raised concerns about surveillance, misuse and public oversight.
• GPU export controls may go remote: U.S. officials are reportedly considering rules targeting Chinese companies that access Nvidia GPUs through overseas cloud providers in countries including Thailand and Singapore.
• Claude Code meets prompt injection: A security researcher demonstrated how a malicious website could influence Claude Code through a summarization request and potentially trigger attacker-controlled commands.
• SpaceX wants more power, faster: SpaceX is building turbine-blade casting capacity in Texas to accelerate gas-powered infrastructure for AI data centers, drawing pollution concerns.
• Nvidia pauses some financing deals: Nvidia halted certain AI cloud revenue-sharing arrangements after employees raised concerns that the structure could attract antitrust scrutiny.
• Uber gets a nearly $1B algorithmic headache: Dutch regulators fined Uber nearly $1B over automated driver suspensions that regulators said lacked sufficient human review.
• Claude Code limits get tweaked: Anthropic announced weekly usage changes coming Sept. 14, prompting criticism from users over how the company characterized the adjustments.
• AI video outruns the clock: Fal demonstrated an “infinite stream” of generated video using Minimax H3 Max, which it says can create five seconds of video in roughly three seconds.

🔧 Tools of the Day

• Grok Bot: xAI’s agent platform added shareable Bots and agentic shopping, including the ability to make purchases through Link using a single-use payment card. The update pushes AI agents further into tasks involving actual transactions.
• H3 Max: Fal’s post-trained video model is tuned for speed and quality, with the company demonstrating generation speeds fast enough to continuously produce video faster than playback consumes it.
• Blotato: The social automation tool connects AI systems including Claude and ChatGPT to automated social DMs, using responses, lead magnets and offers to turn engagement into sales conversations.

Today’s Sources: The Internet, The Neuron, AI Secret, The Rundown AI