Unlocked Today: the price of AI stopped only going down

The cheapest AI on the market raises its prices today, by as much as 1,100 percent. The direction of AI pricing is no longer one way, and the reason is capacity, not competition.

The price floor moved up

DeepSeek, the Chinese lab whose cheap models started the price war, raises its API prices today at 16:00 UTC. The increases run from 50 percent to more than 1,100 percent depending on the model, the token type, and the hour you call it. On its cheaper V4-Flash model, a million output tokens go from $0.28 to $1.32 at peak. Its stated reason is to "allocate resources more reasonably." Read that as: it is out of capacity, and rationing with price.

Pull the other direction. Google shipped Gemini 3.7 Flash at half the launch price of the model it replaced three weeks ago, and xAI's Grok 4.6 matches GPT-5.6 Sol on independent benchmarks at roughly half its price.

Frontier capability keeps getting cheaper. Capacity does not. This week those two curves separated in public.

What this means for you: stop treating this year's price as next year's ceiling. Anything you run on a schedule has a monthly cost you should be able to say out loud. Go find it, then ask what a 3x change would do to you.  → the price change  → the scale of it  → Gemini 3.7 Flash  → Grok 4.6

A neurosurgeon proved a math conjecture he was not qualified to attempt. Jin Shanmu is a neurosurgery resident at Peking Union Medical College Hospital with no advanced mathematics training. Teaching himself the math behind brain-ultrasound research, he locked GPT-5.6 off the web inside ChatGPT Work, started an autonomous run, and walked away. Sixteen hours later it had a proof of Crouzeix's conjecture, open since 2004. Three named experts checked it, including Michel Crouzeix, who posed the problem. It holds. What this means for you: his whole contribution was framing the problem and waiting. You are probably sitting on a question you never posed to a model properly, because you assumed you were not the right person to ask.  → the proof  → the verification

The gap between companies using AI well and everyone else tripled in five months. OpenAI published its own enterprise telemetry: firms in the top 10 percent of AI usage now produce 8.3 times as many output tokens per active user as a typical firm. In January that multiple was 2.6. A Liferay study of 500 US professionals fills in the rest: 54 percent of companies run or pilot AI agents, 25 percent measure the impact against a clear KPI, 24 percent have a written AI policy. The most-cited barrier to getting more value was lack of training, ahead of both cost and security. What this means for you: most companies are already using this and cannot tell you whether it worked. Pick one workflow you handed to AI, and write down what it cost and what it returned last month.  → the telemetry  → the study

Someone hid instructions to AI inside a court filing. On August 6 a Connecticut judge sanctioned a self-represented plaintiff for embedding white three-point-font text in official filings, aimed at any AI reading them. One line told the model to side with him and overturn a clerk's denial. The judge barred him from electronic filing. He called it an audit of whether the court uses AI, then hid more instructions in later filings.

404 Media calls it the first prompt injection attack on a US court caught in the wild. It surfaced the same week researchers documented a four-day AI agent campaign against Taiwan's government that took 85 accounts and 2,500 personnel records, on open-source tooling anyone can download.

Anything your AI reads can try to give it orders: a résumé, an invoice, a vendor's PDF, a web page. If you have an agent touching your inbox, it is reading text written by people who do not work for you.  → the sanction  → the ruling  → Taiwan

Unlocked Today is free. Swept by Nova, our AI news agent; edited down to what matters. Two emails a week, Sunday and Wednesday, and nothing else.

Past editions live in the Vault → go-unlocked.com

Reply anytime. I read them.
Kendall

You're getting this because you subscribed at go-unlocked.com.  Unsubscribe