AI Market Signals: Unpacking the 89X Cost Gap & Model Surge

The performance gap is eight point six percentage points. The price gap is eighty-nine X. Last week's episode was about finding the signals that actually matter. This is the loudest signal in the market today: the cost of near-identical AI performance now varies by a factor of eighty-nine. The market just had a convulsion. On May twenty-ninth, we saw a flood of new models drop all at once. Gemini 3.5 Flash, Qwen3.7 Max, GPT-5.5, and a half-dozen others from Mistral and DeepSeek. It’s a simultaneous leap forward. But the real news isn't just the new models. It's the new realities they create.

First, OpenAI's latest model just solved an eighty-year-old math problem. The Erdős problem 90. This isn't just another benchmark win. This is the first time an AI has autonomously produced a mathematical result that top-tier mathematicians are calling genuinely new. Fields Medallist Timothy Gowers said he'd recommend it for publication in a top journal, no hesitation. This is AI graduating from pattern-matching to producing novel, expert-level insight. Second, the money is chasing the consequences. Cognition, the AI agent company, just raised over one billion dollars in a Series D, at a twenty-six billion dollar valuation.

Their agent, Devin, is now writing eighty-nine percent of their own code. We'll come back to that. Third, a startup called Geordie AI just raised thirty million dollars to build security for AI agents. They map how agents access tools and data, and flag risks in real time. This isn't a feature. It's a new category of security, because the agents are already here. And a final, sobering note. For all this power, Huawei just put GPT-5.5 through a new benchmark called Claw-Anything. It tests complex, real-world digital assistant tasks. The pass rate was thirty-four point five percent.

That’s a hard number to spin. Let's connect the dots. The two most important stories today are the eighty-nine-times price collapse and the billion-dollar funding for Cognition. They seem like opposites, but they are the same story. Let's go back to that eighty-nine times price difference. A new analysis shows DeepSeek V4 Flash hits seventy-nine percent accuracy on a key coding benchmark for just twenty-eight cents per million tokens. Claude Opus 4.7 gets eighty-seven percent accuracy… for twenty-five dollars. The performance difference is marginal. The cost difference is a chasm.

This is happening because the underlying tech—the mixture-of-experts architecture—is getting cheaper to run. And cost-optimized labs, especially out of China, are relentlessly driving the price down. The raw intelligence is becoming a commodity. A utility. Like electricity. So where did Cognition find a billion dollars? They aren't selling electricity. They're selling the machine that runs on it. Their AI agent, Devin, is now pulling in four hundred ninety-two million dollars in annualized revenue. It's being used by Mercedes-Benz, by Goldman Sachs, by the US Army.

A bank in South America, Itaú, is using Devin to automatically resolve seventy percent of its security vulnerabilities. This is the shift. The money is moving up the stack. Away from the raw model, and toward the specialized agent that performs a job. End-to-end. Cognition's twenty-six billion dollar valuation isn't for having a smart model. It's for having an autonomous software engineer that actually ships code. The market just split in two. On one side, a firehose of increasingly powerful and shockingly cheap base models. On the other, a handful of companies proving that the real value is in the agent that can reliably use that intelligence to do a specific, high-value job.

Everyone was watching the race to build a better brain. That race is now over. And it ended in a price war. The game is no longer about who has the smartest model. It's about who has the most effective agent.