On Monday, DeepSeek confirmed that the 75% promotional discount on its V4 Pro API — originally set to expire May 31 — will become the permanent price. The rate drops to one-quarter of the original, effective June 1. No fanfare. No “limited time” asterisk. Just a quiet line in a pricing update: this is the new normal.

On the same day, a project called Reasonix hit the front page of Hacker News. It is a terminal-based coding agent — not built by DeepSeek, but by an independent developer working under the handle “esengine” — engineered specifically around DeepSeek’s prefix-cache architecture. The headline number from a real-world case study: 435 million input tokens in a single day, a 99.82% cache hit rate, and a total bill of roughly $12. The same workload on an uncached model would have cost around $61.

Read those two stories together. The pricing announcement looks like a straightforward competitive move — DeepSeek undercutting OpenAI and Anthropic on cost, again. The Reasonix project looks like a clever community hack. But together they reveal something the billions flowing into model-training startups have obscured: the moat is moving. It is no longer in the weights. It is in the cache.

The 80x Multiplier Nobody Talks About

The economics of large language model inference are not primarily about the cost per token. They are about the cache-hit ratio. When a model processes a long context — a codebase, a conversation history, a set of documents — the expensive work is computing attention over all those tokens. If the system must re-read the entire context for every new query, the cost scales linearly with the context length. But if the key-value cache from the previous turn is preserved and reused, the incremental cost of a new query can be 80 times lower.

This is not a small optimization. It is the difference between a coding agent that costs $61 per day and one that costs $12. The Reasonix case study makes this explicit: the developer ran it for a full workday — 435 million tokens — and the bill came to lunch money. The architecture writeup describes three pillars: cache-first loop, tool-call repair, and cost control. The first pillar is the one that matters. The agent is designed to “leave it running” — to maintain a persistent session where the prefix cache never invalidates, so every new prompt rides on the coattails of everything that came before.

DeepSeek’s API supports this natively. Most competing APIs either do not, or they do so poorly enough that real-world cache-hit ratios hover in the 60-70% range rather than 99%. The price cut, read in this light, is not just a race to the bottom on raw token costs. It is a bet that the layer where money is made — and lost — will be determined by inference architecture, not model architecture.

The Community Built the Agent. DeepSeek Didn’t.

This is the part that should make model companies uncomfortable. Reasonix was not built by DeepSeek. It is an open-source project by a developer who saw the cache architecture, understood its implications, and built the tool that exploits it. DeepSeek provided the infrastructure; the community captured the value.

“The model’s good, but the cache is the product now,” one developer who contributed to the project’s early testing told me in a Slack DM. “If your agent loop can’t keep the prefix stable across turns, you’re burning money no matter how cheap the per-token price is.”

This is not the story the AI industry has been telling itself. The prevailing thesis — the one that justified $80 billion in venture capital flowing into foundation-model companies since 2023 — is that the moat is in the model weights. Train the biggest, smartest model, and you win. Everything else is a wrapper. But wrappers are starting to look a lot like the main event.

Consider what happens if the cache-engineering advantage is durable. The best coding agent is not necessarily the one with the smartest underlying model. It is the one with the agent loop that best exploits the cache architecture of whatever model sits beneath it. And that loop can be built by anyone — a solo developer in a weekend, as Reasonix demonstrates.

The Price Cut Is a Signal, Not a Strategy

DeepSeek’s permanent price cut is worth reading as a strategic signal. When a company makes a promotional discount permanent, it is admitting that the promo price was the real price all along — that the higher “list” price was never going to hold in a competitive market. This is not generosity. It is acknowledgment that the pricing power in foundation models is weaker than anyone wanted to believe.

The interesting question is what DeepSeek does next. If the moat is shifting to inference architecture, the logical move is to build — or acquire — the agent layer. Offer a first-party coding agent that exploits the cache better than any third party can. Capture the value before the community does. But there is a tension here: the thing that makes Reasonix work is that it is open, hackable, and purpose-built by someone who lives in the terminal. A first-party agent from DeepSeek might never match that fit.

The model companies have spent three years arguing that scale is the answer — more parameters, more data, more compute. But Reasonix and the V4 Pro pricing announcement suggest something different: that the next phase of competition will be won by whoever designs the inference stack that wastes the fewest tokens. That is not a model-training problem. It is an engineering problem. And the engineers who solve it may not work for the companies that trained the models.

What Gets Priced In

The market has not yet absorbed this shift. Valuations for foundation-model companies still assume that better models command pricing power indefinitely. But if the cache-hit ratio matters more than the parameter count, that assumption breaks. A model that is 10% smarter but 5x more expensive to run in a real agent loop will lose to a model that is good enough and cheap enough to leave running all day.

This is not a prediction that DeepSeek wins and OpenAI loses. It is an observation that the entire category is being repriced around a variable most investors have not learned to measure. Cache-hit ratios do not appear in earnings calls. They do not appear in benchmark leaderboards. They are not the stuff of keynote announcements. But they are starting to determine who pays what, and to whom.

The price cut takes effect June 1. The agent is already on GitHub. The moat is visible now — it just is not where anyone was looking.

Sources