DeepSeek V4 Flash 3 Cent Test Exposes AI Pricing Deception
A Choice Without a Chooser
Somewhere in a server rack, a model is deciding what to say next. It weighs probabilities, picks a token, and moves on. The machine does not realize it is making a decision. It has no idea that the token carries a price. It will never learn that analysts are watching its output and measuring the value of a thought. The model does not know what anything costs, least of all itself. That is the first thing worth understanding about the new model from DeepSeek: the machine does all the choosing, and none of the calculating is its own.
The calculation belongs to the people who priced it. Their arithmetic has produced the most aggressive number in the industry right now.
The Price That Reads Like a Typo
DeepSeek, the Chinese startup that stunned global markets in early 2025, released its V4-Flash model on Friday. [1] The pricing looks like an error: $0.14 per million input tokens and $0.28 per million output tokens. [1] A token is a fragment of a word, a unit of data that AI systems consume the way engines consume fuel. At these rates, a typical request costs fractions of a cent. The number is so small that it stops feeling like a price. It reads like a rounding error with a decimal point attached.
The small number is precise, and it is intentional. A Reuters report cites research from the firm Artificial Analysis showing that V4-Flash is by far the least expensive to run among all well-known models globally. [2] Anthropic’s Claude Fable 5 costs more than one hundred times as much. [3] That is not a discount. That is a statement.
The statement lands in a market that is still on edge, two years after DeepSeek’s earlier model, R1, triggered a selloff in global technology stocks. Investors suddenly questioned the billions American companies were pouring into artificial intelligence. If a Chinese startup could deliver capable models at a fraction of the cost, what was everyone else paying for? V4-Flash is the follow-up to that question. It is the company doing what it does best: offering ultra-low-cost alternatives that force the rest of the industry to justify its price tags.
The Comparison That Changes the Question
A price tag alone tells us little. The real cost of a model depends on how much work it must do to finish a task. A model with a low headline price can still drain a budget if it needs many steps to produce an answer. A pricey model that answers in one pass can turn out to be cheaper in practice. That is why Artificial Analysis did not stop at the list price. It calculated the average cost per test, a measure that accounts for the actual data a model must process and generate to complete a job.
The results are stark. V4-Flash averages three cents per test. [2] Moonshot AI’s Kimi K3 averages 86 cents. OpenAI’s GPT-5.6 Sol runs at $1.86 per test. Claude Fable 5 checks in at $3.15. The spread between the cheapest and the most expensive offering is wider than the gap between a street stall and a luxury boutique.
The comparison does something more important than display a spread. It shows what is particular about this moment in 2026: a Chinese startup is setting the global price floor for artificial intelligence, and the rest of the world is scrambling to respond. DeepSeek once dominated headlines about Chinese AI development. That era did not last. Domestic rivals besieged the company from every side: Moonshot, MiniMax, Z.AI, ByteDance, and Alibaba all moved into the same territory. Every one of them wants a share of the American and European markets. All of them target businesses looking for cheaper ways to deploy AI at scale. The race is not about who builds the smartest machine. It is about who makes the machine cheap enough that it becomes invisible infrastructure.
The Algebra of Value
The same research firm that measured the price also measured the intelligence. Artificial Analysis combines results from nine benchmarks covering coding, reasoning, and workplace-style assignments into a single Intelligence Index. V4-Flash scores 50 out of 100. [2] That puts it level with Google’s Gemini 3.6 Flash. It sits one point behind Meta’s Muse Spark 1.1 and GLM-5.2 from Z.AI. Moonshot’s Kimi K3 reaches 57. Anthropic’s Claude Opus 5, Claude Fable 5, and OpenAI’s GPT-5.6 all land nine or more points higher.
Here is where the deception begins. The cheap model is not a genius. It is a mid-tier performer with a clearance sticker. Yet the gap in price is absurdly larger than the gap in intelligence. Claude Fable 5 costs more than one hundred times as much per test as V4-Flash, and its lead on the hundred-point scale starts at nine points. If price tracked capability, the expensive model would have to score in the thousands. It does not. We are fooled twice: once by the low price, which makes the model look weaker than it is, and once by the high price, which makes the expensive models look stronger than they are.
Neither number describes what a model actually does. A score of 50 does not tell you how a machine handles a messy conversation, a strange request, or an unfamiliar context. A price of three cents does not tell you what happens once thousands of businesses depend on that model for critical work. The figures are real. They are also incomplete, and incompleteness is a form of deception when it is presented as a summary.
A Market That Runs on Numbers
The artificial intelligence market now runs on figures that look objective and are anything but. Prices are quoted per million tokens, a unit most people will never think about. Scores are quoted on hundred-point scales that flatten wildly different abilities into a single digit. Businesses make procurement decisions based on these digits. They compare, they calculate, they choose. The choice feels rational. It rests on measurements that hide more than they reveal.
DeepSeek is preparing a more powerful model called V4-Pro. [1] It has not announced a release date for that version. On Monday, Alibaba unveiled Qwen3.8-Max, its largest and most capable model to date, not far behind in size from an offering by domestic rival Moonshot launched last month. The launches keep arriving, each one trailing fresh numbers and fresh claims. The competition shows no sign of easing. The numbers multiply, and the confusion grows with them.
The Reversal
And yet the cheapest model on the market makes the strangest final point. V4-Flash scored 50 on the Intelligence Index. Gemini 3.6 Flash scored 50. One of them is by far the least expensive to run among well-known models, according to the same research that produced both numbers. The other belongs to a company that has become synonymous with the AI boom. If the index means anything, the discount model matches the giant. The cheap model is not the one that owes the market an explanation.
The model itself will never know any of this. It remains unaware of its own price. The number 50 means nothing to it. No analysis, no report, no market panic has ever reached it. It simply picks the next token, over and over, without awareness, without ambition, without the faintest idea of the market it has disrupted.
The model is not deceiving us. We are deceiving ourselves. We read intelligence into probability and value into price. We look at a three-cent answer and a $3.15 answer, we see a hundredfold difference in cost and an intelligence gap that begins at nine points, and we conclude that the market is rational. It is not. The one honest participant in this entire industry may be the machine that knows nothing at all, because it is the only one not trying to sell us something.
Sources
1. DeepSeek
3. Anthropic
4. Moonshot AI
5. MiniMax
6. ByteDance
7. Alibaba
8. Google
9. Meta