DeepSeek Strikes Again: 1,114% API Price Hike, Open Harness & Updates

0:00 Thursday morning 11:31 coordinated universal time. Deepseek posts a new flagship model. About 90 minutes later, it posts a coding agent, open- sourced, MIT license, the whole thing, free. And

0:13 in between those two, it posts a new price list. That price list is the part people screenshotted. One cell on it goes up by more than 1100%. And the new rates go live on Sunday afternoon. By

0:24 Friday morning, that free coding agent has passed 55,000 stars on GitHub in one night. That repository was 14 hours old. No launch campaign, no wait list, no keynote. 14 hours. So here's the

0:37 question. The whole day turns on. Why raise your prices and give away your tools on the same afternoon? Both of those turn out to be the same move. And that move is written down in one

0:46 footnote on one page, one sentence under the benchmark table on the model card explaining how those headline scores were measured. Read that sentence and the model, the harness, and the price

0:57 rise stop being three separate stories. They collapse into one. And the piece that most of the coverage got backwards is the model itself. Start there. Because V4 Pro did not get a new brain

1:08 on Thursday. It got a new mouth and a new pair of hands. Start with what shipped. Deepseek V4 Pro finally left preview nearly 4 months after the April version. It is live in the app and on

1:20 the web behind a setting called expert mode and it sits on the API under exactly the same model name it had the day before. Nothing you already wrote has to change. It did not arrive the way

1:30 launches usually arrive. The model showed up on Open Router on Wednesday before Deepseek had published anything about it. Simon Willis writing that day said he had to link to Open Router

1:40 because Deepseek had no obvious announcement page for their new model. The words came after the weights and the weights are right there. The August 13th build sits on hugging face under an MIT

1:50 license, not gated, 92 files, 892 GB of 8-bit weights. Underneath is the same shape as April. 1.6 trillion parameters, 49 billion switched on per token, 384 routed experts, six firing at a time, a

2:07 mixture of experts model with a million token context window trained on more than 32 trillion tokens. Hold on to that 892 GBTE figure because it matters later. The model DeepS is about to

2:19 charge nearly $4 a million output tokens for is one you are free to download and run yourself today for nothing. So what changed between April and August? The model card answers in one line and the

2:31 line is unusually blunt. The August build is in Deepseek's own words built on the preview model structure with a D-Spark speculative decoding module attached. Same architecture, same

2:41 parameter count, a serving optimization bolted on, plus a lot of post- training aimed squarely at agents. DSpark is worth a moment because Deepseek open sourced it back in June. Speculative

2:52 decoding means a fast draft process guesses the next several tokens and the big model checks them in one pass. Deepseek's own measurements put it between 57 and 78% faster per user on V4

3:04 Pro against their previous baseline. That speed now ships inside the checkpoint. Therefore, the second change makes sense. The reasoning effort parameter now has three settings instead

3:14 of a switch. Low, high, and max. Deepseek's guidance is specific. Low for simple tasks, high for daily agent workflows, max for the hard ones. Max will happily emit 384,000 tokens of

3:27 output, which at the new peak rate is a decision with a price tag. The third change is the one that reaches furthest. V4 Pro now speaks the OpenAI responses API natively with what DeepC calls

3:38 one-click codec setup. So you can point OpenAI's own coding agent at a Chinese open weights model by editing a single config line that is distribution dressed as a compatibility note. There is also a

3:50 small strange detail and Simon Willis is the one who caught it. He ran the same drawing prompt at all three reasoning levels and got very different results from each. He says he has not noticed

4:00 that kind of difference from any other model. The effort dial is not a slider on the same answer. It appears to change what the model decides to do. But the strangest part of the launch was where

4:09 the numbers came from. According to Willis, the benchmark table was released into the official DeepSeek Wechathat group, copied and pasted into a Reddit post, which was then deleted and finally

4:19 surfaced as an ASI art table on Hacker News. A trillion dollar industry and the scores traveled by group chat. They are on the official model card now and they are extraordinary. On terminal bench

4:30 2.1, the August build scores 87.9 up from 72.1. On Cyber Gym, 83.3, up from 52.7. And on Deep Sue, a coding agent benchmark, it goes from 12.8 to 62.7. That is a jump of 49.9 points. 49 points

4:50 on the same architecture. So, what actually moved? The answer is note one printed directly under that table. And here is what it says. For the code agent tasks among the public benchmarks above,

5:02 DeepSseek V4 Pro is evaluated with the minimal mode of Deepseek harness as the agent framework using the max reasoning effort level. That one footnote reframes the entire week. So what exactly were

5:14 you buying? The scores were produced by a harness and the harness was open sourced 90 minutes after the model was announced, which means anyone can reproduce the setup. Now look at what

5:23 minimal mode actually is from Deepseek's own documentation. Bash and a file editor. That is the entire tool set. A mode that exists in their words for benchmarking. So the number is a pair.

5:35 It is not this model scores 87.9. It is this model inside this harness in this mode at this effort level scores 87.9. Swap any one of those four and the number moves. That is not a criticism of

5:48 DeepSeek. Deepseek is the one who wrote it down. Which brings us to the giveaway. Deepseek harness version 0.1 developer preview MIT licensed written in Typescript. You install it by typing

6:00 npx at deepseeai/dsh web. It opens a browser interface on your own machine. The repository landed with more than 12,000 commits already in its history, which tells you this was

6:13 not built last month. The design has one idea repeated everywhere. Everything is a plug-in. Models are plugins. Tools are plugins. Skills, sessions, sandboxes, files sims, the agent loop itself,

6:26 theuler, and the interface are all plugins. You can pull any one out and drop a different one in without touching the source. The second idea is that every run is traceable. Everything the

6:37 model sees goes into an appendon event log. the system prompt, the reasoning, every tool call, every result, every context injection tagged by source, resume, fork, search and replay, all

6:49 read from that same stream. So you can rewind an agent run to any step and branch a different decision out of it. And underneath all of it sits a project few people outside its own community had

6:59 heard of, Cordis, a meta framework for spatia temporal composability. Four years old, 1,829 stars. And on the day the harness shipped, a paper formalizing it went up

7:11 alongside plugins you can remove and have every side effect they caused reverted. Hot reloading with proofs. Here is the part that explains the timing. Cortis has one dominant

7:20 contributor, a developer who goes by Shigma with 537 commits to it. His GitHub profile lists his company as Deepseek and the harness team is led by QEN who spent nearly 9 years at Jane

7:33 Street in Hong Kong and joined DeepSseek in March, 5 months before this launch. He is also the repository's top committer with more than 5,000 commits to his name. So, the shape of it becomes

7:44 clear. Deepseek did not grab an open source framework in a hurry. The developer who spent four years building the exact abstraction and agent runtime needs now works there and the runtime

7:54 was built on top of that work in public with a paper and DeepSeek is explicit about why the architecture is open. You can plug in any component you choose including models from other companies

8:05 and it draws the contrast deliberately. American rivals it argues hardcode their products so the harness that produced DeepSeek's own benchmark numbers will run a competitor's model just as

8:15 happily. Two details before we leave the repo. There is no issues tab discussions only on a project with 55,000 stars and no public bug tracker which is a choice and the weight ship without a standard

8:28 chat template. So you get a folder of Python encoding scripts instead. This is a developer preview in the literal sense but the harness is free MIT and it runs against a rivals model as happily as its

8:40 own. Therefore the money has to come from somewhere else. And on Thursday, Deepseek said exactly where. The same afternoon it gave away the rapper, it repriced the thing inside it. Now, the

8:50 pricing move did not come out of nowhere. On the 1st of August, one single tool, Open Code, put roughly 8 trillion tokens through DeepSeek V4 Flash in one day. 5 trillion of those

9:01 were free trials. 3 trillion were paid, 8 trillion tokens in a day through one tool. And for the week before that, V4 Flash was the single most used model on open router on Earth at 7 22 trillion

9:13 tokens. Chinese models took the top four slots on that board. Deepseek had won the price war so completely that the prize was a cue because the compute underneath is finite. Analysts put

9:24 Deepseek's fleet at roughly 20,000 H100 equivalent GPUs, most of which arrived in the previous couple of months. That is a serious cluster. It is not a cluster that absorbs 8 trillion tokens a

9:35 day at 14 cents a million while 5/8s of them are free trials. So, watch the clock. On the 30th of July, OpenAI cut its cheaper models prices by up to 80%. On the 31st, Deepseek shipped V4 Flash

9:49 at 14 cents in and 28 cents out, undercutting it. On the 4th of August, the service started degrading. On the 6th, Deepseek warned that a significant increase was coming with no numbers. On

10:01 the 13th, the numbers arrived, and they are steep, but not uniformly steep, which is the part the headlines flattened. The smallest rise on the table is about 52% V4 Pro Fresh Input

10:12 off peak. The largest is 1114% V4 Pro cached input at peak. Output tokens land in between at roughly 4 and 1/2 times the old rate. The change that actually hurts is the one the headline

10:25 skipped. Under the old prices, reading a token from cash cost 120th of reading it fresh. Under the new prices, it costs exactly 130th. The cash discount did not disappear. It shrank by a factor of

10:39 four. If your agent replays a long context every turn, that single line is your bill. Then there is the shape of the day. Peak hours are 1 to 4 and 6 to 10. Coordinated universal time and off

10:50 peak is half price. Convert that to Beijing time and the pattern is obvious. 9 to 12 and 26. Chinese office hours. 7 hours out of 24 cost double. And the discount is a request to run your batch

11:03 jobs while Hangjo sleeps. Deepseek's stated reason is to allocate resources more reasonably, which is corporate for we are out of GPUs. But there is a second reason on the record. The company

11:14 is in talks at a valuation near $71 billion and preparing paperwork for a mainland listing. Loss leader pricing is a growth story and a hard thing to put into a perspectus. So the question that

11:25 actually matters is Deepseek still the cheap option? Yes. At peak, a million output tokens from V4 Pro cost $3.96. Claude Fable 5 is 50. GPT 5.6 Saul is 30. Kimmy K3 is 15. Gemini 3.1 Pro is

11:42 12. Off- peak Deepseek is 198. It is still by a distance the cheapest serious model you can rent. But the gap moved and the direction is what counts. Before Thursday, Deepseek output was 57 times

11:55 cheaper than Fable 5. After Sunday, at peak, it is 12 1/2 times cheaper. An artificial analysis, the only independent measurement published on this build, now calls it somewhat

12:06 expensive when compared to other openweight models of similar size. That sentence about DeepS would have been unthinkable in April. So, here is where I land. The harness is the win, and it

12:17 is not close. Deepseek gave away the layer every other lab keeps welded shut under MIT in Typescript with an install command you can type from memory and it will run a rivals model as happily as

12:28 its own. The model is very good and it is not Frontier 53 on the independent index against Claude Opus 5's 63 and on Terminal Bench the road Deepseek led its own announcement with it comes third

12:41 behind Kimmy K3 and Fable 5. Deepseek published the two numbers that beat it. So the villain in this story was not a company. It is the habit of reading a score as a property of a model when the

12:53 model card says plainly that it is a property of a pair. And Deepseek deserves the credit the coverage did not give them because naming your harness in a footnote and then shipping that

13:02 harness is more disclosure than this industry usually manages. It landed on the same day as the bill, so it got buried under the invoice. Which leaves the question worth sitting with. Look at

13:11 whatever agent you are running right now. How much of what it does is the model you pay for?