Anthropic just told everybody Claude Sonnet 5's intro pricing is going away on September 1. DeepSeek fired back with a model that costs next to nothing. And the guy running Palantir stood up and said he doesn't trust the AI labs. That's your Tuesday in AI.
Claude Sonnet 5's Cheap Pricing Is Almost Over
Anthropic set a deadline. Come September 1, Claude Sonnet 5 goes from $2 per million tokens to $3 per million. That's a 50 percent jump. On top of that, a tokenizer change means the same text can eat up to 35 percent more tokens than before. So you're not just paying a higher rate, you're burning more tokens to say the same thing.
Why it matters: if you built a product on Sonnet 5 assuming today's math, go redo your math. Margins on AI features are already thin for most startups. A hike like this can turn a profitable feature into a loss leader overnight.
Robert's take: I get why they're doing it, compute isn't free and intro pricing was always going to end. But burying a 35 percent token increase inside a price hike announcement is the kind of thing that makes builders nervous about trusting these vendors long term. Read the fine print before September.
DeepSeek Undercuts Everybody at 14 Cents
While Anthropic raises prices, DeepSeek dropped V4 Flash at $0.14 per million tokens. That's not a typo. It's roughly a fourteenth of what Sonnet 5 will cost after September.
Why it matters: this is the pattern we've seen since DeepSeek showed up. Frontier labs push capability and charge a premium for it. Chinese labs chase that same capability a few months later at a fraction of the price. If your business logic runs on the cheapest model that's good enough, you've got another option now, and it's dirt cheap.
Robert's take: cheap and good enough beats expensive and slightly better for most real world use cases. I'd be testing V4 Flash against whatever I'm running today before I renew any contract.
Palantir's CEO Doesn't Trust the AI Labs
Alex Karp said out loud what a lot of enterprise buyers think privately: the frontier AI labs aren't trustworthy enough for enterprise deployment. He said it right after Palantir posted a 93 percent year over year revenue jump to $1.94 billion, with US commercial revenue up 149 percent.
Why it matters: Karp isn't some outsider complaining. He runs a company that sells AI software to governments and Fortune 500s for a living, and business is booming. When someone in that seat says the labs themselves are the risk, that's worth paying attention to, especially if you're the one signing procurement contracts.
Robert's take: Karp has incentive to talk his own book here, Palantir wants to be the trusted layer between you and the model. But the underlying point stands. Model providers keep changing pricing, changing terms, and changing models out from under you. Enterprises want stability more than they want the newest benchmark win, and right now stability is in short supply.