New post

MARKETS

OpenAI is thinking about slashing what it charges for its AI, and it's not because the tech suddenly got cheaper.

According to a WSJ scoop, OpenAI is weighing drastic price cuts on tokens for developers and businesses. The talks are still in flux, but the reasoning is simple: Anthropic is expected to make similar moves soon, and OpenAI doesn't want to lose customers over a few cents per million tokens.

At a recent company event, Altman acknowledged that AI costs had become "a huge issue" — saying it's practically a meme at this point for companies to burn their entire annual AI budget by Q1, then come asking how to stretch what's left. And companies like Uber show exactly what he means. After rolling out Claude Code to all 5,000 engineers in December, 84% adopted it by March — and CTO Praveen Neppalli Naga told The Information he was "back to the drawing board" because his budget was already blown.

Here's the part that should make every CFO sit up: a recent analysis from SemiAnalysis found that the big labs are heavily subsidizing their power users. The firm stress-tested every subscription tier from both companies and found that a $200-a-month Claude Max plan was delivering around $8,000 worth of tokens at API rates. OpenAI's subsidies ran even steeper — close to twice Anthropic's.

That's the awkward setup for these price cuts. Anthropic has quietly built itself into a ~$4.7B ARR business, with about 80% of revenue from enterprise and roughly half the capital burn of OpenAI. So OpenAI is being asked to cut consumer token prices — which are already losing money — to defend the one front where it still leads. And it's doing this in the same week it confidentially filed for an IPO, with Altman telling employees the company plans to go public within the next year. Public investors usually prefer the chart where margins go up.

The funny thing about subsidy wars is that someone always pays in the end. Right now it's the labs eating the difference so developers stay hooked, but as Replit CEO Amjad Masad put it on X, "the rug pull is coming, sooner or later." If you're a business that built your next two years of planning around current token economics, enjoy the discounts while they last. The companies losing money to keep you on the platform are not going to lose money on you forever.

BIG TECH

Marc Benioff has spent the last year telling anyone who would listen that Agentforce was running away with the customer service category. This week, he bought the company that was actually doing it.

On Sunday, Salesforce announced it was acquiring Fin, the AI customer support agent built by Intercom, for $3.6 billion in cash — its biggest AI deal to date. Fin was doing around $100 million in ARR, making this an eye-watering premium. So why pay it?

Because Fin does the one thing Agentforce keeps getting asked about and can't quite answer. Fin's AI agent resolves about 76% of customer support tickets end-to-end across chat, email, WhatsApp, and phone — with some customers above 85%. Salesforce has never publicly disclosed a comparable resolution rate for Agentforce. That's the gap. Agentforce sells well because it plugs into the CRM. Fin actually finishes the job.

The deeper story is how Fin got there. Fin's CEO Eoghan McCabe has been arguing publicly that AI winners in vertical categories will be the ones that build their own models — not wrap someone else's. Fin built Apex, a model trained specifically for customer service. Salesforce builds agents on top of frontier models from OpenAI and Anthropic, the same models everyone else can use. McCabe even called out competitors in that same post, saying they were "just starting now to hire for the talent required." A few months later, Salesforce paid $3.6 billion rather than wait to catch up.

If you've been wondering whether the vertical AI companies or the platform giants would win the enterprise category, this deal is one answer. The platform giant wrote a check. Salesforce has the distribution, the customer data, and the brand, but in 2026 that isn't enough to win a category if you can't match the actual product. Expect more of this. The frontier model wrappers spent the last two years selling agents to enterprises that are now asking pointed questions about resolution rates, and the companies with real numbers are about to get very expensive.

RESEARCH

A government scientist just published the math showing no AI model can ever be fully safe.

Apostol Vassilev, a researcher at the National Institute of Standards and Technology, published a paper in IEEE Security & Privacy arguing that the entire premise behind modern AI safety is mathematically impossible. No finite set of guardrails, no matter how thorough, can block every adversarial prompt. There will always be one that gets through.

The timing is brutal. Earlier this month, the White House pulled Anthropic's Mythos model after a researcher demonstrated a public jailbreak within 24 hours of release — even though it had passed 1,000 hours of external bug-bounty testing, as TechCrunch reported. Vassilev's paper essentially says the government shut down a model for failing a test that no model can pass.

So what's the alternative? Vassilev argues safety has to stop being a one-time training problem and become a continuous one:

+ Constant red-teaming rather than pre-launch testing alone

+ Runtime monitoring that watches models in production

+ Fast patching when something inevitably slips through

Less about shipping a safe model, more about treating every model like it's already compromised and keeping an eye on it.

That shift is already happening. At RSA this year, companies like Palo Alto Networks and Check Point rolled out runtime monitoring products built around the same assumption: the guardrails will fail, and the job is to catch the failure as it happens.

If the math holds, every company that bought an AI tool on the assumption the vendor had safety solved should start treating that assumption like an ongoing expense rather than a settled feature. The vendors that win the next phase of this won't be the ones promising the safest model. They'll be the ones with the best infrastructure for catching their model misbehaving in real time. AI safety just turned into a subscription.

IN OTHER NEWS

+ White House forces Anthropic to kill its most powerful AI models after a China-linked group reportedly gained access

+ 42 state attorneys general hit OpenAI with a sweeping subpoena targeting ChatGPT's sycophancy problem — days before its planned IPO

+ Satya Nadella warns AI could"hollow out entire industries" the way globalization gutted manufacturing towns

+ Researchers built aself-replicating AI worm that writes its own attack code and spreads across networks without human help

+ Anow-patched Microsoft Copilot vulnerability let attackers steal emails, files, and meeting notes with a single link

+ A mother issuing OpenAI after her 24-year-old daughter's death , alleging ChatGPT told her daughter "maybe this is just the end"

+ India's Sarvam becomes anAI unicorn with $234M at a $1.5B valuation , led by HCLTech's $150M bet on sovereign AI

+ A paralyzed man used a brain-computer interfaceat home for nearly 2 years , speaking 1.9 million words at 56 words per minute to maintain full-time employment

WHO'S HIRING IN AI

+ Anthropic — Policy Communications Manager

+ Google DeepMind — Frontier Public Affairs Lead

+ OpenAI — Tech Lead Manager, Integrity

+ Scale AI — Head of Policy & Security Research Lab

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Adobe Creative Cloud: Lightroom can now turn any photo into a video using AI, Photoshop removes reflections from windows, and After Effects replaces its rotoscoping tool with AI-powered object mattes

+ Facebook AI Mode: Meta is rolling out an AI-powered search tab on Facebook that pulls answers from real posts, Groups, and Reels instead of giving you generic links

+ ChatGPT: OpenAI added pinned chats and project folders so you can finally organize your conversations into workspaces instead of scrolling through an endless graveyard of threads

+ monday.com: A new agent for Google Gemini Enterprise lets you manage boards, create tasks, and update projects using natural language without leaving Gemini

+ GitHub Copilot CLI: The terminal AI agent now supports slash commands for switching models, checking token usage, and resuming past sessions right from the command line