Does the price of intelligence matter?
Until recently, outside of being a fan of the idea of open-weight models, I didn’t tend to be that interested. Their value proposition was to be cheap, but the value proposition of the frontier models was to be truly intelligent.
Until recently the only way you could pay for reasoning and intelligence, was to pay people, and typically a lot. The more specialized knowledge and the better at reasoning the higher the rate and the harder it is to even find someone available with the skillset.
In comparison to the time of a financial analyst, a software developer or BDR, combined with their limited availability, the ability to pay an API provider for some of this time was not at all challenged by cost of the tokens involved, but by whether the model was intelligent enough to actually add comparable value.
This is still mostly the case. Really strong intelligence is typically much more important to successful output than cost. The main limiting factor on most tasks that we could theoretically apply AI to solve, is still mainly that the agents are not capable of fully doing the task in a reliable way. The main challenge to doing more with AI is still mainly a question of intelligence, not price.
DeepSeek V4 Flash is the first time a model fundamentally makes me feel that this time it’s different. It’s not a frontier model. Fable, GPT 5.6 Sol are clearly ahead and Opus 5 is stronger as well, but working with it on coding tasks, or tasks like publishing this article (I typically publish any of my articles directly from Agent Runners), feels very similar to working with a model like Opus 4.8.
You need to instruct it well, but it’s thorough, it can take on hard problems, it can run long autonomous sessions and it is strong enough that you can use it to work on real world production software. But it is an order of magnitude cheaper than any of the comparable models from OpenAI/Anthropic.
Other models have been almost good enough and somewhat cheaper. That typically ends up just not mattering. Actually good enough is what counts.
But now I’m starting to see a whole set of problems where DeepSeek is absolutely good enough, and so cheap that I’ll use AI for it even if it was just not viable before.
It feels like we’re now suddenly at the tipping point where we are seeing a drastic split between “absolutely most intelligent” models, where limit is still intelligence, not price, and where we’ll continue to see a brutal battle between labs to deliver the SOTA model at any given time.
And then a completely different race to make intelligence ubiquitous and almost free, that’ll be dominated by open models and lead to massive growth in the inference platforms supporting them.
This is a new era opening up for AI.
Typesetting and editorial work by DeepSeek V4 Flash, Cover design by Kimi K3