In the API, the model specified is “deepseek-flash”; today, you are accessing version V4.1 of the Flash 1017 model. Two and a half months ago, when using the same name, you were actually accessing a different model. During this period, the name of this interface has been changed three times, and the corresponding model information in the documentation has also been updated accordingly; however, the code used by those invoking this API has not been modified in the slightest.
When comparing it with the Gemini 4 Hydrogen version, this particular parameter needs to be considered as the most important factor. The “actual performance tests of DeepSeek Flash” that you may have seen online are likely not referring to today’s latest version of this product.
On the other hand, the name “Hydrogen” was introduced quite recently—only on November 12th—and therefore all the publicly available data related to this model are consistent, making the information relatively more accurate. However, the amount of data available is limited, and only one or two third-party organizations have conducted retests.
The two models were combined because last night, DeepSeek made its “full suite of APIs” available to the public. Functions such as online searches, advanced analytical searches, file input, link reading, and voice recognition are all now integrated into the V4.1 version of their platform. On Google’s end, the Hydrogen model was also incorporated into their global search system on the same day. Both companies are moving in the direction of providing users with more “built-in tools” within their respective models, during the same week.
Several different versions have been used under the name deepseek-flash.
The version “1017” was launched on October 17th; the numbers in its name simply represent the date of release. Prior to this, an earlier V4.1 Flash version had also been available under this same interface name. When Meituan released the performance report for LongCat-2.5 on November 1st, the DeepSeek column only listed “V4.1-Flash” without indicating whether it referred to version 1017 or not. As a result, readers of the report had no way of distinguishing which specific version was being compared.
This has two practical consequences for you.
Firstly, the speed tests and price estimates conducted online in September cannot be directly applied to today’s situation. Secondly, the prompts used during those online tests might have been changed due to the use of a different algorithm on a certain morning. Therefore, regression testing should be carried out on a monthly basis, rather than quarterly.
Hydrogen doesn’t have this problem, but it suffers from another type of misnomer: Google refers to it as the successor to Gemini 3.8 Flash, yet its price is actually twice that of Gemini 3.8 Flash. Being called “Flash” already means it is being sold at a mid-range price.
Those extra switches that appeared today…
The tool added by DeepSeek to V4.1 Flash can be used in three interfaces: Messages, Responses, and Chat Completions. The server handles the tool calls and context integration itself. The supported file formats include PDF, Office documents, and images. Functions such as Search, Fetch, OCR, ASR, and TTS can also be invoked separately. The official team has also released a front-end demo on GitHub, which allows users to import historical conversations.
Create a “customer service assistant that can search the internet and read PDF files.” In the past, three or four different services were required to achieve this functionality; now, just one key combination along with a few additional parameters is enough.
Hydrogen’s approach to toolization takes a different route. Instead of providing a ready-made “package” at the API level, Google has integrated it into its own products: in search, it allows a single query to be broken down into multiple sub-topics for simultaneous searching, with the results page directly displaying comparison tables and useful tools; within Antigravity, there is an Orchestrator that enables Carbon or Argon to act as the primary intelligent agents, assigning tasks to several Hydrogen-based subordinate agents. The difference is quite clear. DeepSeek offers its capabilities to developers so that they can build solutions on their own, while Google uses Hydrogen as a “tool” within its own products. If you plan to build an application from scratch, DeepSeek’s framework can be used immediately; if you are already working within Google’s ecosystem, Hydrogen has already taken care of the integration for you.
37 minutes and 2.5 times
1017: Yeah, it was clearly a case of the vehicle flipping over while moving.
On October 18th, some domestic media successfully replicated the official demonstration example: When set to the default “High” configuration, it took 37 minutes to build a 3D model of the Forbidden City, but the resulting image was just an isolated palace; for the same task, GPT-6.1 Sol completed it in just 7 minutes. In another test where it was asked to create a racing game, the game itself managed to run after 42 minutes, but the steering wheel was set in reverse – turning left actually made the character move right.
The main selling point of Hydrogen is its speed. According to Google’s figures, when using the high-performance inference mode by default, the speed at which tasks are completed is 3.8 times that of Flash. Although this figure has not been independently verified by third parties, the general consensus in the community supports this claim: Hydrogen indeed allows users to complete tasks without having to wait for long periods of time.
Therefore, the biggest difference between these two models does not lie in the scores themselves, but rather in how long one has to wait after pressing the Enter key. For tasks that are processed in batches, offline, or in the middle of the night, the slight delay caused by model 1017 is practically negligible. However, when a person sits in front of the screen and makes repeated changes to the code, producing a semi-finished product in just 37 minutes means that it is almost unusable.
The difference between fast and slow lies in the way they are used.
Whether it’s five times or fifty times – it all depends on the cache hit rate.
First, let’s list the prices. For Hydrogen, it costs $1.5 per million tokens input and generates $7.5 in output. For the 1017 service: during peak hours, the cost is $0.30 per million tokens input and $1.20 per million tokens output; during off-peak hours, it’s $0.15 per million tokens input and $0.60 per million tokens output. When cache hits are taken into account, the costs during peak hours are further reduced to $0.006 per million tokens input and $0.003 per million tokens output; during off-peak hours, they remain at $0.15 and $0.60 respectively. Just looking at the listed prices, during peak hours, using the 1017 service is 5 times cheaper for input and 6.25 times cheaper for output; during off-peak hours, the costs double, becoming 10 times and 12.5 times cheaper respectively.
What truly creates a significant difference is caching. Take the most common operation performed by programming agents as an example: with a context containing 100,000 tokens, the output consists of 2,000 tokens.
Hydrogen is not considered part of the caching system; the cost per request is 0.165 US dollars. When 10,175 requests do not hit the cache, the total cost is approximately 0.032 US dollars. However, if all 100,000 tokens are cached successfully, the cost per request drops to just 0.003 US dollars.
There is a difference of more than fifty times between the two values.
There is a figure that I cannot find: the price per unit for caching and reading data related to Hydrogen. Google has not made this information publicly available yet. If Hydrogen also offers a discounted caching price, around 10%, then this difference would be reduced to just over ten times. Until Google releases this information, it is safer to assume that no discounts are being applied.
Another factor affecting pricing lies within Google itself. Currently, Argon is available at a discounted price of $2/10, but this will increase to $4/20 starting from December 1st. Hydrogen scored 51 points in the Artificial Analysis comprehensive index, which is one point higher than Argon’s score of 50. After December, many daily tasks that were previously performed using Argon will be shifted to Hydrogen. Due to its lower price within Google’s product lineup, it seems like a more cost-effective option. However, when compared with DeepSeek, the overall cost situation remains essentially the same.
The problem with hydrogen doesn’t lie in its speed.
Google claims that the performance of Terminal-Bench 4.1 has increased from 18.2% to 57%, while third-party tests on version 4.0 showed approximately 54%. The scores themselves are quite good.
What makes me hesitate are the hallucinations it generates. Some evaluations have shown that the proportion of fabricated content in its responses is approximately twice that of Argon. It’s already common for fast models to make mistakes, but now Hydrogen appears at the very top of Google search results, and its answers often come with tables and conclusions right away – leaving users with fewer reasons to verify the information. It’s fast… and it generates fake content just as quickly.
There is another obstacle that domestic developers simply cannot bypass: the Gemini API cannot be directly connected from the Chinese mainland; proxy servers or overseas cloud services must be used instead. This is not the fault of the Gemini API itself, but it certainly adds to the costs associated with using it.
The problem with 1017 isn’t its low price at all.
A common opinion on the Internet is that “1017 models represent affordability, but their performance is average.” I don’t quite agree with this view; it misidentifies the weaknesses of the 1017 series.
In terms of performance, it scored 77.0 points in DeepSWE v1.1, outperforming Claude Opus 5’s score of 74.0. On OpenRouter, it consistently ranked first in weekly usage rates – real developers who actually used it gave it their approval through their actual usage data.
Its real weaknesses lie in stability and terminal capabilities. On Terminal-Bench 4.0, DeepSeek reported a score of only 42.3%, which is approximately ten points lower than Hydrogen’s third-party scores – and this was still just a self-reported value. On AA-Omniscience, the correct answer rate was 43%, meaning that more than half of the knowledge-based questions were answered incorrectly. Coupled with the issue of interface name changes mentioned earlier, it is clear that DeepSeek is more suitable for tasks where users can simply try again if they make mistakes or wait patiently if the process takes a long time to complete.
Another risk relates to platform promotions. On Tencent WorkBuddy, the 50% discount promotion for points related to the “1017” program ended on November 5th without any prior notice. As a result, users had their accounts charged at the full price for an additional three days, and only half of the amount was refunded afterwards. For those who use third-party platforms to manage their “1017” points, it is advisable to keep track of the expiration dates yourself.
End up engaging in several specific types of activities.
Tens of thousands of operations per day, involving bulk extraction and rewriting with long contexts— coupled with the addition of caching in version 1017—make it virtually unbeatable. The search and file-input features now available also enable it to easily handle another batch of requests that previously would have required developing custom crawlers.
In interactive programming where users wait for results in front of a screen, the fact that Hydrogen offers 2.5 times faster performance is far more important than its price – provided that you are willing to accept the occasional errors or bugs that may occur during its operation. Please verify this key conclusion yourself before making any decisions.
For products designed for overseas users and already running on Google Cloud, what Hydrogen eliminates is the hassle associated with setting up connections.
For domestic teams with limited budgets and those willing to accept slower progress, simply start with option 1017 and conduct a regression analysis once a month.
Relevant comparisons: “Gemini 4 Hydrogen vs. Claude Sonnet 5.5: 57% and 70.6% – these are absolutely not the same set of test questions.” “Kimi K3.1 Pro vs. DeepSeek V4.1 Flash 1017: The score difference between each question is as much as eight times; the extra money spent was only used to purchase the terminal programming module.” “LongCat-2.5, GLM-5.3, DeepSeek V4.1 Flash: On Meituan’s report card with 31 lines of results, all the scores in our own column were actually self-test scores.”