Last week the story was that DeepSeek had built a death zone. Any model that cost more or performed worse than V4-Pro was, per Bloomberg, simply obviated—the cheapest credible frontier engine had swallowed the middle of the market whole.
This week the same lab raised output prices more than fourfold, then quietly deleted its own claim that the new model had "significantly enhanced agent capabilities."
Hold those two facts next to each other, because the gap between them is the whole story. A company does not usually delete a boast about its flagship in the same week it charges more for it.
A company does not usually delete a boast about its flagship in the same week it charges more for it.
The numbers first, because every developer already ran them. From 16 August, per Bloomberg, a million output tokens through DeepSeek-V4-Pro costs $3.96 at peak, up from $0.87. V4-Flash reportedly climbs from $0.28 to $1.32, per Bloomberg.
Off-peak rates are half the new peak, which produces the detail that actually stings. Off-peak V4-Pro at $1.98 is more than double the old undiscounted price of $0.87.
A team that shifts every workload to the quiet hours still pays over twice what it paid at lunchtime last week. For V4-Pro output, the discount floor is now reportedly higher than the old ceiling.
The instinct is to read this as a cost report—compute got expensive, so prices rose. The engineering says otherwise.
V4-Pro is a 1.6-trillion-parameter mixture-of-experts system with 49 billion parameters active per token, and DeepSeek says its attention design cuts the compute needed for a single token to 27% of the previous generation. That efficiency is precisely what made the old prices possible in the first place.
So a fourfold rise reads as a decision about margin rather than a report about costs. Nothing in the hardware got four times harder; the company simply decided to charge more.
The reason it decided to is not hidden. Bloomberg puts the increase directly against an IPO, reporting DeepSeek is preparing to list as soon as this year at a valuation around $71bn. Founder Liang Wenfeng now has to show investors a path to profit, not just a benchmark chart.
Loss-leading is a private-company tactic. It works when your only job is to take share. It stops working the moment you have to explain your gross margin to a public market.
But re-pricing the engine is the smaller half of the move. The larger half shipped the same week.
On Thursday DeepSeek released a developer preview of DeepSeek Harness—the scaffolding that wraps a model so an agent can read files, edit code, browse the web, and keep going until a task is done. The model is the engine; the harness is the car.
That is exactly the layer Anthropic monetizes as Claude Code, and it is where the money in agentic coding actually sits. DeepSeek had been telegraphing the ambition for days, standing up a public "DeepSeek Harness Team" account and posting jobs describing the goal as turning its models into cutting-edge agentic products.
The design choice is the tell. DeepSeek's harness is deliberately open—users can plug in any component, including models from rival companies—and the company framed that against American tools it says hard-code their own model underneath.
Read the strategy, not the marketing. An open harness that runs anyone's model is a bid to own the workspace instead of the engine.
If developers live inside your tool, the model underneath becomes a swappable part—and the part you would most like them to swap to is your own. That is why the price hike and the harness are one move, not two.
Owning the harness is how you stop needing to loss-lead. Once the workspace is the product, the model can be re-priced toward margin because it is no longer the thing keeping developers in the door.
Which is why deleting that "significantly enhanced agent capabilities" line matters more than a tidy edit should. The statement marked V4-Pro leaving a preview that ran nearly four months, and by Thursday afternoon it was gone, per the South China Morning Post, with no explanation.
You do not usually retract a claim about your flagship in the same week you quadruple its price—unless the flagship is quietly being demoted from headline act to interchangeable component.
The developer reaction underneath the strategy is the crack in it. Early reactions to the 0813 build left developers underwhelmed on general capability and unhappy about the price, per the SCMP, though researchers were impressed in narrower areas like cybersecurity.
The benchmarks are vendor-reported and unreplicated. DeepSeek's own card puts V4-Pro at 80.6% on SWE-bench Verified, level with Gemini 3.1 Pro and a hair behind Claude Opus 4.6 at 80.8%—but 67.9% on Terminal Bench 2.0 against GPT-5.4's 75.1%, and 37.7% on Humanity's Last Exam against Gemini 3.1 Pro's 44.4%.
Roughly level on coding, behind elsewhere. That is the model whose price just rose fourfold.
There is a related asterisk worth keeping. Together AI's analysis reportedly showed V4-Pro hitting 88.5% pass@4 at $0.24 per task—but a reader noted the chart showed 62.8% pass@1. The cheap-per-task story depends on the model getting several attempts; the first-attempt number is not frontier.
At $0.24 a task you can afford retries. But retries only stay cheap while the per-token price is low, and that is exactly the variable DeepSeek just moved.
Now the fragility, which is the point. The hike does not end the death zone—it narrows it. At $3.96 per million output tokens, DeepSeek still sits well below Moonshot's Kimi K3 at $15 and Anthropic's Fable 5 at $50.
Cheap is relative, and DeepSeek is still cheap. What ended is the claim that Chinese inference is effectively free.
The problem is the shape of the bet. DeepSeek just raised prices fourfold on a model developers found underwhelming, deleted its own boast about it, and asked those same developers to move into a harness so it can monetize their work.
Every one of those users now has a reason to look at the newly vacated floor. If Kimi or GLM holds its price while DeepSeek climbs, the rational move for a cost-sensitive team is to route to whoever hasn't blinked—and the open harness DeepSeek built will happily run those rivals' models instead.
That is the trap in an open architecture. The same design that lets DeepSeek swap its own model in also lets developers swap it out.
So the strategy only survives on one condition: the harness has to get adopted outside China before a cheaper rival undercuts the floor DeepSeek just abandoned. Own the workspace and the price hike is a margin story with an IPO attached.
Fail to own it, and the price hike is just a discount removed—handing the death zone to whoever is willing to keep living in it.