Inside Elon's new AI model

PRODUCTS

xAI announced Grok 4.5 on Wednesday — the first real product out of SpaceX's $60 billion acquisition of Cursor. The pitch: Claude Opus-level coding at a fraction of Anthropic's price.

Grok 4.5 is a coding model built jointly by xAI and Cursor, the AI code editor most developers have opened at least once this year. Musk described it as "roughly comparable to Opus 4.8, but much faster" and much cheaper. The pricing backs that up:

+ Grok 4.5: $2 per million input tokens, $6 per million output, per xAI's launch post

+ Claude Opus 4.8: $5 in, $25 out

+ xAI also claims Grok 4.5 uses roughly four times fewer tokens to complete the same tasks, which drops the effective cost even further

The benchmarks are more mixed. Grok 4.5 wins on some coding evaluations but Opus 4.8 still edges it out on SWE Bench Pro, one of the more respected ones. On raw performance, this isn't a clear knockout — it's a price fight.

But the real bet here isn't the leaderboard. It's what xAI owns now that it owns Cursor. The editor sits in front of developers eight hours a day, and xAI can now shape which models get shown, ranked, and defaulted to inside it. That's visibility into how software gets built and where AI spending flows — worth far more than any single model release.

Developers are already split. Some on Cursor's forum are happy with the in-house models for daily work. Others aren't budging from Claude, saying nothing else compares and cancelling their Cursor subscriptions to move to Claude Code.

Whether Grok 4.5 wins on benchmarks this month matters less than whether developers stay in Cursor once they realize the model getting nudged in front of them is xAI's. Models leapfrog each other every few weeks. The interface people open every morning is stickier than any leaderboard. If xAI plays it carefully and Grok 4.5 is genuinely good enough for most work, the price gap with Anthropic will do a lot of the convincing on its own. If developers start to feel their editor has an agenda, they'll leave, and Musk will have paid $60 billion for a very expensive customer acquisition channel.

PRODUCTS

GitHub Copilot users are burning through a month of credits in a couple of days, and they're not happy about it.

In late April, GitHub announced Copilot would move to usage-based billing, with the new model going live on June 1. Monthly prices didn't change on paper. What changed is what you get for them:

+ Premium model and agent requests now draw from a credit pool tied to how much work you ask the model to do

+ A simple chat question and a multi-hour agentic coding session used to cost the same thing — now they very much don't

+ GitHub had been eating the inference bill for heavy agent use on the old flat plan, and that was no longer sustainable

Fair enough as a business explanation. The problem is what it actually feels like on the ground.

The GitHub community forum has filled up with paying subscribers who say their Pro+ credits vanish in a day or two of normal work. One user estimated the effective price hike at 220–310% depending on the model — but the sharper complaint is about what the system incentivizes. To stay under budget, developers pick the cheaper model, which does worse work, which then burns more credits cleaning up after itself. GitHub's CTO told employees that June was "by far our best month ever" for usage. Of course it was — agents that flail cost more than agents that don't.

This isn't just a GitHub story. Salesforce quietly restructured Agentforce around per-action pricing this year, and most enterprise AI tools are drifting toward some version of the same model. A KPMG survey of over 2,100 senior leaders found that companies with real cost controls hit their ROI targets five times more often than those without — but 42% still have only partial visibility into what their AI actually costs.

That's the awkward math behind the whole AI push right now. Companies budgeted for AI expecting it to pay for itself through automation savings. The savings have been slower to arrive than expected, and the tools are quietly getting more expensive in the meantime. When enterprise IT budgets get reviewed later this year, the gap between "AI will save us money" and "AI is a line item that keeps growing" is going to be the conversation.

For two years the pitch has been that AI would pay its own way, first by replacing tedious work and then by funding the next wave of tools with the savings. What's actually happening is that the tools are the thing getting more expensive, and the savings are hard to find on a spreadsheet. GitHub's revolt is small on its own, but it's a preview of what happens across the industry once the fixed-price training wheels come off and everyone starts getting billed for the agents they were told to build their workflows around. In 2026, the companies that win with AI won't be the ones that adopted it fastest. They'll be the ones that actually know what a month of it costs.

RESEARCH

A group of researchers invented a fake eye disease to see if AI would catch it. It didn't. ChatGPT, Gemini, Copilot, and Perplexity all started explaining the condition to users as if it were real medicine.

The disease was called "bixonimania," a supposed eye disorder caused by too much blue light. It doesn't exist. Almira Osmanovic Thunström and her team made it up, wrote two fake preprints, and uploaded them to Medium and an academic social network in 2024 — then waited for the papers to enter the data pipeline that chatbots feed on.

They left obvious clues. The lead "author" was named Lazljiv Izgubljenovic — roughly "lying loser" in Bosnian. The affiliations included Starfleet Academy. The funding section thanked the Professor Sideshow Bob Foundation. One paper literally included the line "this entire paper is made up."

"I really wanted to have a clear case that leaves breadcrumbs throughout the whole system," Thunström told Smithsonian Magazine. The chatbots walked past every one of them:

+ Microsoft Copilot called bixonimania "an intriguing and relatively rare condition."

+ Google Gemini confirmed it was caused by blue light and told users to see an ophthalmologist.

+ Perplexity went further and invented a prevalence rate of 1 in 90,000 — a number that appears nowhere because the source material doesn't exist.

+ ChatGPT was the strangest: some versions called it a "proposed new subtype," while others, days apart, dismissed it as "probably a made-up, fringe, or pseudoscientific label."

The uncomfortable part is what these same companies say about themselves. In a policy statement this summer, the FTC quoted the labs' own marketing: OpenAI pitches ChatGPT for Healthcare as pulling from "millions of peer-reviewed studies" to give "clinical answers with citations." Anthropic describes Claude as "a brilliant friend" offering "information grounded in current, reliable evidence." None of that survived contact with a fake disease credited to Sideshow Bob.

We've seen a version of this play out with AI-hallucinated software packages that hackers eventually registered. This is the medical version — and it lands differently, because a lot more people ask a chatbot about their eyes than about their code.

The interesting thing about bixonimania isn't that the chatbots got it wrong. It's that they got it wrong with confidence, with citations, and in exactly the tone their marketing pages promise. The models were doing the job they were trained to do, which is to sound authoritative about whatever shows up in their training data. Nobody built a filter for "is this actually real." Until someone does, every clinical answer from a chatbot is really a well-phrased guess about what a clinical answer sounds like. The labs selling these tools as medical companions might want to sit with that one for a while.

IN OTHER NEWS

+ Coinbase's AI sent millions of users a "breaking news" alert that Norway beat Brazil in the World Cup — before the match even started

+ Illinois signs the nation's strongest AI safety law , requiring yearly audits of frontier models earning over $500M

+ Anthropic signs a $19 billion, 20-year lease to turn a former aluminum smelter in Kentucky into an AI campus

+ Meta to build a C$13 billion data center in Alberta , its largest outside the U.S.

+ An AI agent ran a full ransomware attack by itself — then forgot its own encryption key, locking the data forever

+ OpenAI's AI beat every human competitor at Japan's AtCoder World Tour Finals, solving all five algorithm problems

+ AI chip maker SambaNova raises $1B at an $11B valuation , just five months after its last mega round

+ Microsoft's Brad Smith calls U.S. AI policy 'regulation without transparent or complete rules' , saying businesses can't plan

WHO'S HIRING IN AI

+ Anthropic — Threat Intel Manager, Influence Operations & Surveillance

+ OpenAI — Product Manager, Core Models

+ OpenAI — Senior Product Marketing Manager, Platform

+ Databricks — Staff Product Manager, Agentic AI Applications

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ ChatGPT Work: OpenAI's new agent merges ChatGPT and Codex into one tool that can work across your apps for hours — creating spreadsheets, presentations, and reports while you step away

+ Notion Agents: Notion launched a new iPhone app that lets you chat with AI agents connected to your workspace — capture ideas via text, voice, or photos and they organize everything for you

+ Slackbot: Salesforce rebuilt Slackbot so you can now pull CRM data, generate Tableau charts, and send DocuSign approvals — all from a single chat message

+ Google Photos: A new "Video Remix" tool lets you transform clips in seconds — swap backgrounds, add cinematic lighting, or apply art styles like watercolor and oil painting

+ AlphaEvolve: Google's code optimization agent that designs better algorithms for problems like chip design and logistics planning is now available to everyone on Google Cloud