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

SpaceX confirmed its orbital Starmind data centers will run on Nvidia's Vera Rubin NVL72 racks, with Elon Musk targeting the first space-based racks by late 2027. Each rack packs 72 chips working as a single computer, with Nvidia claiming up to 25 times the compute of its older H100.

Musk says the space version of the hardware is simpler, cheaper, denser, and lighter than typical server gear, tuned for orbit's radiation and heat instead of a data center floor. Analysts peg orbital compute at more than 4 times the cost of ground compute today, a gap Musk claims will close within a few years.

Read Nvidia's announcement.

The bigger signal: 👉 Sam Altman called space-based data centers "ridiculous" earlier this year. With ground-based buildouts now facing real public opposition (a recent poll found three-quarters of Americans don't want one nearby), the timeline for orbit stopped looking like a stunt and started looking like a backup plan.

Every sales leader knows the feeling. You walk into a pipeline review with a number you believe in, and twenty minutes later, you're defending every line item to a CEO who just wants to know what's actually going to close.

HubSpot Sales Hub ends that conversation. Every deal, every rep's activity, and every buyer signal are all in one place and updated automatically. So your forecast is built on what's actually happening. And when you present that number, you can stand behind it.

Get a demo

OpenAI dropped API and credit pricing on GPT-5.6 Sol, its top-performing model, by more than 20% for the next three months. Sol now costs $4 per million input tokens and $20 per million output tokens, down from $5 and $30, undercutting Anthropic's Opus 5 ($5 input, $25 output) and Fable 5 ($10 input, $50 output) on both ends.

The cut follows last month's price drops on GPT-5.6 Terra and GPT-5.6 Luna, of 20% and 80% respectively, and lands as enterprises get pickier about what intelligence actually costs. Data from payments platform Ramp shows Anthropic's flagship Fable 5 makes up just 6% of business token spend, while OpenAI's Sol accounts for 25% of tokens used by OpenAI customers on the same platform.

Read OpenAI's pricing announcement.

The bigger signal: 👉 Enterprises are already voting with their token spend for efficiency over raw capability, and OpenAI cutting its best model's price says it read that shift before Anthropic did. Watch whether Anthropic follows with cuts of its own.

Anthropic made Claude Mythos 5, its most capable model, available through Claude Security, letting Enterprise customers scan their own codebases for vulnerabilities and get suggested fixes. Until now, Mythos 5 access ran exclusively through Project Glasswing, Anthropic's cybersecurity partnership program.

The rollout has been deliberately slow. Anthropic has treated Mythos 5 as its most safety-sensitive release, and opening it to Claude Security customers marks the first time the model reaches enterprises outside that partnership.

Read about Claude Security.

The bigger signal: 👉 Anthropic keeps its most capable models on a short leash longer than any other lab, and widening Mythos 5's access to paying enterprises signals it has enough confidence in the model's safety profile to let it loose on customers' actual production code.

Route by task, not by default Most API calls don't need your most expensive model. Reserve the frontier tier for the step that actually requires deep reasoning, and hand routing, classification, and simple extraction to a smaller, faster model. A common pattern: a cheap model decides what kind of request just came in, then only escalates to the frontier model when the task genuinely calls for it.

Watch total cost, not sticker price A model listed as 70%+ cheaper can still cost more to run once retries, longer completions, and extra back-and-forth get counted. Track cost per completed task, not cost per token, before assuming a cheaper model actually saved you money.

Check whether a subscription beats the API If your usage is moderate and steady, a flat-rate subscription plan can undercut metered API pricing once you add up actual monthly spend. Compare both before committing to a routing setup built entirely around the API.

Consider a local model for high-volume simple tasks For tasks like intent classification, tagging, or short extraction that run constantly, a local open-weight model can cut costs to near zero once it's set up, at the cost of some setup time and infrastructure.

The bigger signal: 👉 The "route to the cheapest model that can do the job" pattern is becoming a standard part of production AI architecture. Build it in from the start and every future price cut turns into pure savings instead of a rewrite.

A few tools quietly worth exploring:

You’ve read the AI breakdowns. Watched the million AI explainers. Nodded along to the hot takes.

And then opened a blank doc and had no idea what to actually do.

The Shift is a newsletter built to help you in moments like this. Every tool covered works in the real world, and every prompt in the 1000+ library solves something you’ll actually hit. 

3,000+ vetted tools. 1000+ tested prompts. Daily newsletter.

And right now, 3 subscribers win a free 1-year Claude Pro subscription. One click to enter.

Subscribe to The Shift

This issue is about who controls the layer everyone else builds on. Nvidia's chips are headed to orbit before ground-based data centers finish losing their popularity contest. OpenAI cut its best model's price days after enterprise customers quietly stopped paying for Anthropic's most expensive one. Anthropic answered by finally trusting its most guarded model with customers' actual production code. Different companies, same fight: control the infrastructure, and you control what gets built on top of it.

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

See you next issue.