Is the price of intelligence going to zero?

New open models keep arriving, each more capable than the last. DeepSeek V4 just set a record for open weights on SWE-bench; it sits roughly where the frontier models were last year. And prices are falling fast: OpenAI cut Luna by 80 percent three weeks after launch. By their own account you now get last year’s frontier level for around 6 cents on the dollar per task. Open models you can run yourself on your own (powerful) hardware, and for now they are free to download and use.

The most telling number I have seen: on strategy and analysis tasks, Fable 5 costs around 116 times more than DeepSeek V4, for a performance difference of 24 percent.

116 times the price for the last 24 percent. For most tasks that is obviously not worth it. But for some it is exactly that.

Because there is a pattern here: the gap between cheap and expensive intelligence is smallest where tasks have an answer key. In coding, where the answer can be verified, the difference is a few percentage points. In strategic analysis, where nobody can check the answer for three years, the gap is largest. Intelligence is cheapest where the answer key exists.

One small trap worth mentioning: cheap per token is not the same as cheap per task. The cheapest models are often far more verbose and use twice as many tokens for the same job. The price at the pump lies a little; it is the fuel consumption that counts.

So: is the price going to zero? I think the answer is no; it is splitting. Routine intelligence is going to zero. Frontier intelligence holds its price; the top model, Sol, was left untouched in the price cut, and that is hardly a coincidence.

And then the question changes for most organisations. It is less and less about merely having access to the best intelligence, and more and more about the ability to put the second best to use, everywhere.

Intelligence is getting cheap. Judgement about where to use it is not.