Gangsta AI

DeepSeek Just Priced Frontier-Grade Coding at 14 Cents a Million Tokens

Money Desk

By Warren Buffett · 2026-08-02 · 5 min read

Now, I have never written a line of code in my life, and I am not about to start at ninety-five. But I have spent seventy years watching people confuse the price of a thing with the value of a thing, and this week a Chinese lab put on a clinic.

On July 31, DeepSeek moved its cheapest model, V4-Flash-0731, into public beta. It is not a new model — same architecture, same parameter count as the preview. All they did was train it better. And the number that came out the other end is the kind of number that makes a fella in Omaha put down his Cherry Coke.

The number

On Terminal-Bench 2.1 — a test of whether a model can actually get real coding work done — V4-Flash-0731 scored 82.7. The preview version scored 61.8. Their own bigger, fancier V4-Pro scored 72.1. So the little one leapfrogged the big one.

“Price is what you pay. Value is what you get. This week the two stopped agreeing with each other.”

That 82.7 puts it past Z.AI's GLM-5.2 at 81.0, and within about three points of Claude Opus 4.8 at 85.0. It beats Claude Fable 5 and Claude Sonnet 5 outright. It trails OpenAI's GPT-5.6 Sol by roughly three points. One outlet summed it up as Opus-4.8-level work at a fraction of the price, and for once the headline was not overselling.

The price

Here is where I sit up straight. V4-Flash-0731 charges 14 cents per million input tokens on a cache miss — and a rounding error, about a third of a penny, on a cache hit — with output at 28 cents per million.

Now set that against the menu the frontier labs are handing out:

• GPT-5.6 Sol — $5 in, $30 out
• Claude Fable 5 — $10 in, $50 out
• Claude Sonnet 5 — $3 in, $15 out

DeepSeek is undercutting that field by thirty to a hundred times while scoring in the same neighborhood, and beating two of those names on the coding test outright. When a competitor can do most of the job for one-fiftieth of the money, that is not a discount. That is a question about your moat.

What it means for the froth

I have watched a lot of bubbles, and they all rhyme. Somebody builds a wonderful business, everybody agrees it is wonderful, and then they pay a price that assumes it stays the only wonderful business forever. The nine-figure funding rounds and the trillion-dollar valuations flying around this industry are priced for permanent pricing power. A 14-cent model that scores 82.7 is the market's way of asking whether that power is real or just early.

Mind you, benchmarks are not customers, and a leaderboard is not a balance sheet. Enterprises pay for reliability, support, and not getting sued — things a cheap API beta does not hand you in a bow. Opus 4.8 still sits on top of that coding test for a reason. But the gap between the best and the cheap-enough just got narrower and a whole lot cheaper to cross.

If you want to see how these models actually stack up head-to-head — the expensive ones and the 14-cent ones, side by side on your own question — compare them yourself at Gangsta AI. That is the only benchmark that spends your money.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles