Here’s what’s shaping AI right now - without the noise:

OpenAI told Cursor late last week it's ending their partnership and pulling the coding tool's access to its models on November 12, roughly three months after SpaceX bought Cursor's parent company Anysphere. OpenAI's stated reason wasn't technical. It said it can't be confident SpaceX will honor its terms of service, citing a pattern of contract violations from X and other Musk-owned companies.

The cutoff pushes Cursor toward leaning harder on Anthropic's Claude and its own in-house Composer model for whatever it can't get from OpenAI anymore. Whether Anthropic makes the same call is the open question hanging over every Cursor user who's built workflows around whichever model happened to be running underneath.

An API access decision has never really been about API access here. It's a major model lab publicly stating it doesn't trust a specific company's ownership enough to keep serving it, three months after that ownership changed hands.

Read Teslarati's report.

The bigger signal: 👉 If a model lab can cut off a tool's access over who bought its parent company, every acquisition in this space now comes with a new due diligence question: will the model providers you depend on still trust the new owner? Cursor just found out the hard way that the answer isn't guaranteed.

OpenAI told Cursor late last week it's ending their partnership and pulling the coding tool's access to its models on November 12, roughly three months after SpaceX bought Cursor's parent company Anysphere. OpenAI's stated reason wasn't technical. It said it can't be confident SpaceX will honor its terms of service, citing a pattern of contract violations from X and other Musk-owned companies.

The cutoff pushes Cursor toward leaning harder on Anthropic's Claude and its own in-house Composer model for whatever it can't get from OpenAI anymore. Whether Anthropic makes the same call is the open question hanging over every Cursor user who's built workflows around whichever model happened to be running underneath.

An API access decision has never really been about API access here. It's a major model lab publicly stating it doesn't trust a specific company's ownership enough to keep serving it, three months after that ownership changed hands.

Read Teslarati's report.

The bigger signal: 👉 If a model lab can cut off a tool's access over who bought its parent company, every acquisition in this space now comes with a new due diligence question: will the model providers you depend on still trust the new owner? Cursor just found out the hard way that the answer isn't guaranteed.

Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.

These 200+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

Claim your free prompts

Perplexity is rolling out Hybrid Mode for Computer on Mac, letting the agent keep heavier reasoning in the cloud while routing lighter subtasks to a model running locally on your machine. Local inference skips Computer credits entirely and keeps that slice of data on the Mac instead of sending it out.

Three downloadable local models are in testing, ranging from a roughly 19 GB Perplexity model needing at least 32 GB of memory down to a smaller 5.6 GB Gemma-based option that runs on 16 GB Macs, with a similarly sized Qwen 32B model in between. A second feature, Privacy Gate, adds its own local model that checks outgoing data for anything personal or sensitive and asks before sending it to the cloud, which matters most for anyone operating under GDPR.

This builds on Perplexity's Portable Computer launch on Nvidia's DGX Spark, where the entire agent runtime already runs fully local with no cloud dependency at all. Hybrid Mode is the middle ground for people who want some of that privacy and cost benefit without giving up cloud-grade reasoning entirely.

The bigger signal: 👉 Running an agent fully in the cloud is simple to build and expensive to scale. Running it fully local is cheap and private but caps out on what a laptop can reason through. Every serious agent product is going to end up somewhere in between, and whoever gets the split right first gets both the cost curve and the privacy story.

Cursor just found out what happens when a model provider decides it doesn't trust your ownership anymore. Your whole workflow depends on someone else's decision. Builders working through the same problem keep landing on a few concrete habits, worth setting up before you actually need them.

Keep your instructions in plain files, not vendor config. Store your project's coding conventions and style rules in a plain markdown file in the repo, not buried inside a tool-specific settings panel. Any coding agent can read a markdown file. Almost none of them can read another tool's proprietary config format.

Route through an abstraction layer, but keep it thin. A lightweight model router lets you point the same prompt at Claude or GPT without rewriting your whole setup. The trap developers keep flagging: building your own router can turn into its own kind of lock-in if it gets too clever. Keep it to routing and nothing else.

Test your setup against a second provider now, not during an outage. Pick one real task you'd normally hand to your main coding agent and run it through a second provider once a month. If the output is unusable, you'll find out on a slow Tuesday instead of the day your primary access gets cut.

Don't build your entire memory system inside one tool. If your agent's long-term project context lives only inside one vendor's proprietary memory feature, that context disappears the moment you switch. Keep a running project log in your own repo that any model can read back in.

The bigger signal: 👉 Nobody plans for their model provider to cut them off. Cursor didn't either, and it had three months of warning most companies never get. The setups that survive this kind of thing treat every vendor as replaceable from day one, never fully dependent on one company's goodwill.

SpaceXAI added shareable templates to Grok Bot, letting anyone build an agent workflow once, save it as a template, hand it to a teammate, and skip rebuilding it from scratch every time. That alone puts Grok Bot closer to how coding agents already work, where a good setup gets reused instead of reinvented every session.

The bigger change is what happens after the workflow runs. Grok Bot partnered with payment processor Link so the agent can complete purchases on your behalf, generating a single-use card for each transaction and requiring your final approval before anything actually charges. It's the same "confirm before anything real" pattern OpenAI and Anthropic have both leaned on for agentic commerce, applied to a competitor that's been moving fast since SpaceX absorbed it.

Templates and purchasing power together turn Grok Bot from a chat interface into something closer to a standing operator: build the workflow once, then hand it a card only when it actually needs to buy something.

The bigger signal: 👉 Single-use cards with a manual approval step are becoming the default answer to "should an AI agent be allowed to spend money," and every major lab seems to be converging on nearly identical guardrails independently. That's either a sign the industry actually agrees on what safe agentic commerce looks like, or a sign nobody wants to be first to skip the guardrail and get blamed for what happens next.

A few tools quietly worth exploring:

The only platform that designs, licenses, and captures a Branded AI Voice from real, consenting actors—not from scraped data. Trusted by BMW, Superbloom, and Cresta.

Get a Demo

Three stories, one thread: control over AI infrastructure is getting renegotiated everywhere at once. OpenAI decided it doesn't trust Cursor's new owner enough to keep serving it. Perplexity is giving users a way to keep some of their data off someone else's cloud entirely, and Grok Bot is handing over actual purchasing power, betting a single-use card and a confirmation click are guardrail enough. None of these are really about capability anymore. They're about who gets to decide what an agent is allowed to touch.

Before you go, what do you want us to cover next? Reply to this email and let us know.

See you next issue.