From 68 RMB to 500 RMB, we've tested Doubao Pro for you. We feel that the product's design direction meets our needs, but its capabilities still need improvement.

Last week, the long-rumored paid version of Doubao finally arrived—Doubao Pro was officially launched, including three tiers: a continuous monthly subscription for 68 RMB, and fixed subscriptions for 200 RMB and 500 RMB, corresponding to Standard, Enhanced, and Premium packages, respectively.

According to official statements, Doubao Pro is based on the latest Doubao 2.1 series large model and introduces new services for complex work and productivity scenarios. It will provide higher usage limits for productivity scenarios and integrate the Doubao 2.1 Pro model's work task mode.

So, how does Doubao Pro actually perform? Is it worth paying for?

We invited three users from different industries with varying usage habits to conduct first-hand tests. Their needs were vastly different. However, after testing, we found some good points as well as some common issues.

Here are a few conclusions:

First, and most importantly, the usage limit runs out too quickly! It's like opening a blind box—you never know if the limit will suddenly be exceeded after completing a task.

Second, the work mode is decent, with good analysis and planning, but bugs often occur during execution. Its ability to act as a 'mouthpiece' is acceptable, but its 'hands-on' ability is lacking, requiring repeated debugging, which still poses a barrier for ordinary users.

Third, complex industry research analyses still suffer from hallucinations, image recognition analysis remains inaccurate, and coding capabilities are usable, placing it slightly above passing overall.

Fourth, the audio transcription function is indeed excellent, especially in handling English and professional terminology, outperforming many specialized tools. This single feature has the potential to attract heavy users to pay.

01 Standard Version | 68 RMB/month: Just configuring tasks consumes most of my limit

Tester: Doubao user, primarily for tasks such as gathering information, writing, and image generation

As a Doubao user, I mainly used it for search, writing, and image generation tasks in my previous work. Is there a significant difference in effectiveness between the paid and free versions in these scenarios? Tests show that in image generation tasks, both modes allow Doubao to accurately understand requirements and produce error-free images. However, the paid version delivers more stable details, higher clarity, and watermark-free commercial use.

But can Doubao's work version handle some essential tasks that I've only dreamed of but never accomplished?

Scenario 1: Analyze and clean up my critically full C drive

In Doubao Pro's work task mode, there is a feature that supports local computer and browser operations, which reminded me of my frequently full C drive. If Doubao could first help me deeply analyze space usage, clearly distinguish which files can be safely deleted, which need confirmation, and which should never be touched, and then gradually assist in cleaning up, that would be fantastic.

So, I gave it a try. Upon receiving the instruction, Doubao immediately began scanning and analyzing my C drive. After about 20 minutes and over 30 steps of thinking and execution, it finally produced a 'C Drive Space Analysis Report.'

Overall, the results were quite good. In addition to the content I requested, the report also provided cleanup priorities and expected outcomes. It even uncovered four screen recording videos occupying 10 GB of space that had been hidden for nearly three years.

However, when the task moved from summary analysis to cleanup execution, problems began to arise.

Doubao performed decently when handling 'brainless deletable' files. But when it came to deleting the four screen recording files, it prompted that they could not be deleted due to being in use by a program. Subsequent cleanup tasks were interrupted, and it prematurely provided a final cleanup summary.

Even weirder, at this point, the C drive space increased instead of decreasing, going from over 3 GB remaining to just over 2 GB. It suggested restarting the computer. After several attempts, I ultimately had to manually clean up the space. When executing deletable installation packages, another bug occurred—it claimed they were deleted, but memory usage remained unchanged. When I showed it a screenshot, it even misidentified 'red' as 'blue' and told me the red light issue had been resolved.

In summary, Doubao's work mode can indeed operate the computer, but its execution capabilities still have many bugs, making manual operation faster for me.

Scenario 2: Podcast and video transcription

The second task stemmed from my lack of time to watch videos or listen to long podcasts daily. Could Doubao generate verbatim transcripts and summaries directly from links?

Tests revealed that the content extracted by Doubao was slightly more detailed than the podcast's text introduction but extremely simplified compared to the entire podcast content. So, it was unclear whether it truly parsed the podcast link or to what extent. Below are two images showing my request and its response:

However, Doubao's real-time audio transcription function worked very well. The summaries were clear and well-formatted, with additional features like key quote summaries, verbatim transcripts, and key decisions.

Most importantly, its transcription of English and professional terminology was nearly error-free. Knowing that my iFlytek recorder, which cost over a thousand RMB, often fails to correctly transcribe popular professional terms like 'ChatGPT,' 'Agent,' and 'embodied AI,' requiring additional time for verification, I was impressed that Doubao needed minimal corrections.

A fellow user also felt the same way after experiencing it and even considered abandoning their iFlytek account, which had accumulated years of material, to pay for Doubao instead.

According to Doubao's audio transcription usage limits, the free version allows a daily cumulative duration of up to 90 minutes, with a single session limit of 30 minutes. For light users, the free version suffices. However, heavy users will find the Pro version more suitable, offering a total limit for advanced features like audio summaries five times that of the free version (equivalent to 450 minutes daily). However, in practice, '1 minute of recording ≠ 1 minute deducted from the limit,' as each step consumes resources, including real-time transcription, voice separation, timestamp annotation, and full-text semantic understanding—each invoking large model inference, resulting in an actual duration far less than 450 minutes.

Scenario 3: Timed AI morning news briefing

The third essential scenario was an AI morning briefing. Could I have Doubao automatically gather global AI hot topics from the past 12 hours every morning at 9 and save them as a Word document on my desktop?

Upon receiving the instruction, Doubao created an 'AI Morning Briefing' folder on my desktop. However, the next morning, it didn't push any briefing to me. It wasn't until I inquired that I discovered a timing misalignment issue—it thought it was still the previous afternoon, leaving the task in a waiting-to-trigger state.

About 7-8 minutes after I corrected it, Doubao produced a decent AI daily report, with verified information mostly accurate. It also informed me that the task would run daily from then on. However, the next day, the timed push still failed to trigger automatically, again due to timing misalignment.

To resolve this, I switched to using Doubao's 'Skills' section, creating a 'Daily Auto-Running AI Morning Briefing Skill' and confirming the timed push repeatedly. Still, it failed to trigger. Could it not support timed tasks? I sought help from a friend using the 500 RMB version and found they did receive timely pushes.

Interestingly, on July 1st, the day after I temporarily gave up debugging, the AI morning briefing unexpectedly triggered automatically. After several more attempts, the timed task now runs successfully.

Summary:

After testing several scenarios, my overall impression is that Doubao Pro performs well in 'thinking' tasks like information gathering and summary analysis, but its execution capabilities need improvement. More critically, the Standard version's limit is somewhat insufficient—configuring tasks alone consumed most of it. Tasks I planned to test, such as PPT generation and automatically capturing bidding data into spreadsheets, couldn't run before the limit was exhausted, leaving me to wait until the next cycle, a week later, for new limits.

As for whether I'm willing to pay for it? At this stage, it hasn't fully convinced me. I'll continue using the free version, especially in scenarios where I'm already proficient, as it suffices. For more complex tasks, the Pro version's 'cost-effectiveness' hasn't yet met my standards.

02 Enhanced Version | 200 RMB/month: From 'What can it do?' to 'Am I not using it right?'

Tester: Former corporate culture worker who used Doubao intensively daily

I previously worked in corporate culture, with a highly varied workload. I used AI, especially Doubao, intensively every day for interviews, writing various copy, planning schemes, coordination, and review, ranging from group message copy to project proposals and OKR filling. Leaving AI behind, I wouldn't know how to function in the workplace.

My company also strongly encouraged AI use. A month before I left, our team was tasked with creating an AI cultural atmosphere within the company. My team leader excitedly said, 'Our goal is to make all colleagues who haven't used Xia yet feel like they'll be left behind by the times.' At that time, Xia was very popular, and my team leader, having tried it first, was anxious about being replaced by AI. Honestly, I was relieved to be leaving, sparing me from this new project.

After leaving my job, I moved to Yunnan and still habitually use Doubao, but more for life's trivial matters. Just the other day, while walking my dog, I didn't notice it lick an ant. A few minutes later, its legs went weak, and its tongue turned white, resembling shock. While hailing a taxi to the pet hospital, I asked Doubao what to do. It immediately told me how to proceed, and I rinsed the dog's mouth as suggested. Later, I discovered the ant was a red imported fire ant, an invasive species that can cause allergic reactions in dogs and even humans. I shared this information with our community property management, along with Doubao's management suggestions.

Scenario 1: Xiaohongshu note and vlog generation and publishing

After subscribing to Doubao Pro, being unemployed, I wanted it to help me manage a Xiaohongshu account for my dog. I had opened the account over six months ago to document its growth but had been updating it casually without a clear positioning (positioning) or persona (persona), amassing only 90 followers in six months.

I hoped Doubao Pro could post Xiaohongshu notes daily based on the pictures and videos I provided, enabling regular account operation, freeing my hands, clarifying the account's positioning, and ideally growing it.

I assigned it several tasks: first, create an account operation plan; second, post a note daily; third, edit a vlog of my dog from the provided pictures and videos.

It excelled at the first task, as writing plans was its forte, detailing positioning, content direction, and update rhythm convincingly. Unfortunately, it failed at the latter two tasks.

First, auto-publishing. After struggling all afternoon, I realized it couldn't log into my Xiaohongshu account due to anti-scraping mechanisms—it couldn't bypass CAPTCHAs. Most frustratingly, it didn't admit its limitation upfront but pretended to execute, only revealing its inability when I inquired why nothing had been posted after a while. After over ten attempts, I gave up: instead of wrestling with it over login issues, I could simply click publish myself.

Next, editing the vlog. After finalizing the script, I provided it with 60 photos and 9 videos, asking it to select shots based on the copy. The result was unwatchable: the screen suddenly switched from vertical to horizontal, several seconds of black screens appeared mid-video, and the promised 25-second video came out at either 20 or 14 seconds. After 11 revisions, none were usable. Repeatedly, I felt it would be quicker to edit it myself than waste time with it.

During this process, I reflected on my approach. Maybe providing a more detailed script, specifying which shot to use for each scene and precise feedback on what to change and when, would enable it to succeed. I realized that behind every great video lies a team of meticulous editors, even for AI-generated videos. As someone unwilling to invest much effort, it was unsurprising that I couldn't obtain a satisfactory video.

Scenario 2: WeChat Read book card creation and updates

I also asked it to organize my WeChat Read reading history and notes into a systematic personal reading library—since I always forget what I've read and feel my reading categories are fixed, desiring a precipitate ( precipitate - accumulating/systematizing) system.

Although Doubao initially claimed it couldn't be done and even suggested using Claude instead, it eventually organized my past reading preferences and created book cards based on my pain points.

These book cards not only organized the books I've read, including highlighted content and notes, but also linked related books and compiled our discussions.

I also asked it to localize them, so even if I stopped paying for Doubao, a double-click would sync updates to my WeChat Read highlights and notes.

However, just as I thought everything was set, I discovered it wasn't automatically updating the book notes. After pointing this out and being assured it would auto-pop a webpage post-update, it still failed. When I inquired again, it showed my limit had been exhausted, requiring me to wait until July 4th for further use, merely two and a half days into my Enhanced version experience.

Still obsessed with the WeChat Read book cards, I asked the free version why the automation wasn't working. It troubleshooted and insisted the program was fine, blaming my missing node environment. At this point, I sought help from my programmer boyfriend. After reviewing my conversation with Doubao, he was shocked that I'd given it my cookie, a security-related item. I verified this with Doubao, which then said:

So, I asked again:

Previously, Doubao had said:

Although I was an ignorant user, Doubao hadn't fulfilled its duty to remind and inform me at that time.

Reviewing the so-called automated book card organization program Doubao wrote, my boyfriend found it wasn't automated at all but generated a few static webpages instead.

Simultaneously, I tried the free version—installing the official WeChat Read skill on IMA (Tencent's official tool). It could also analyze my reading preferences and data, but when asking it to organize my read books' notes, it cited insufficient computing power.

Realizing Doubao had also failed, I tried the free WorkBuddy (Tencent's Agent) with my boyfriend's help, installing the WeChat Read skill for note organization and automation tasks, which currently fulfills about 60-70% of my expectations.

Summary:

Am I willing to keep paying for it? No, I'm not.

Using the paid version requires having numerous scenarios and tasks for it to handle. For me, the free version's copy polishing and general Q&A suffice. Timed reminders and auto-publishing seem like created needs—upon reflection, the functions I attempted can be achieved step-by-step with the free version. Striving for automation resulted in more time spent on setup, feeling counterproductive.

Second, using the paid version has a learning curve and threshold (threshold). It's not as simple as typing a few words in the chatbox to get things done. You need to write prompts effectively, design workflows, and understand its capabilities and limitations. My quick exhaustion of the limit also stemmed from improper usage. These past few days, I even felt anxious, constantly searching 'What else can Doubao Pro do?' When it performed poorly, my first reaction wasn't 'This function is bad' but 'Did I write the prompt wrong? Am I not using it right?'

Last but not least, safety is a critical concern. When using Doubao Agent, it does not prompt safety risks for tasks it cannot complete. Instead, it suggested a method that required me to provide cookies. However, most ordinary users are often unable to discern whether they should provide the information being requested.

Of course, this is just my personal opinion. If you have a significant amount of repetitive copywriting, planning, and data processing tasks at work every day, the professional version might indeed save you a lot of time.

03 Advanced Version | 500 RMB/month vs. 20 USD Codex: Who Wins?

Tester: Former employee of a major company, currently working in OPC, heavy daily AI user

AI tools compared: Doubao Professional (500 RMB/month, Pro 2.1) vs. Codex (20 USD/month, GPT-5.5)

Since late last year, I've been paying for AI services and now use ChatGPT and Codex daily. Their roles are clearly divided: Codex handles project-based tasks like software development, in-depth research, and stock analysis, while ChatGPT is used for daily search queries and writing novels, showcasing stronger language capabilities.

Currently, my highest-frequency use case is stock market review. I've set up an automated project that generates a daily review every evening—covering data, price changes, capital flows, sector popularity, and even position recommendations. For news, Codex sends me a daily digest of major U.S. company news, eliminating the need to browse multiple websites myself.

I'm also exploring some experimental tasks, such as breaking down marketing work into reusable skills, using AI for image-based social media accounts by combining trending topics with visuals, and attempting to have AI write stylized novels. Another idea is to use AI for information gathering and organization to provide recommendations for parents unsure about which extracurricular sports training suits their children, as there's currently no review system like Dianping for such institutions—this is worth trying.

Additionally, I occasionally use some free products like DeepSeek, which has decent Chinese language capabilities. I rarely use Zhipu. I used Doubao less before because its early versions had severe hallucination issues, making it hard to tell whether the answers were true or false without verification. In that case, it was better not to use it at all.

Now, Doubao has introduced a professional version, which, according to its description, is evolving into a productivity tool. This is already a consensus among global large model companies—pure To C Q&A doesn't generate revenue, so everyone is shifting towards productivity tools like office work and programming.

How to test its capabilities? I selected three tasks to evaluate three key areas, running both Doubao Professional and Codex through them to see the results.

Let's start with the conclusion.

Doubao Professional's overall performance falls between 60-65 points—usable but not yet good. Codex (not even the highest version, using GPT-5.5) scored between 75-90 points, so the gap remains significant.

Scenario 1: Industry Research Brief—More Information, Less Accuracy

The first task was to generate an industry research brief on 'China's AI Office Agent Market Opportunities in 2026.' This is one of the most common scenarios in daily office work, primarily testing the AI tool's ability to gather, organize, refine, and summarize information.

Doubao's Performance: The content appeared extensive and comprehensive but fell apart under scrutiny. Its ability to summarize and refine was notably lacking, with a strong sense of information overload and several incorrect pieces mixed in. When asked to verify errors or distortions in the report, Doubao listed at least six major categories of issues. Clearly, from a serious industry research perspective, Doubao's performance was not rigorous enough—if I still need to verify every detail, it's hard to justify paying for such an 'office productivity' tool.

Codex's Performance: It excelled in rigor and summarization skills. However, a clear issue remained—overseas models still have defects in capturing Chinese-language materials.

Score: Doubao Professional 60 vs. Codex 75

Scenario 2: Product Prototype Recognition and Optimization—Visible but Unclear

Doubao particularly highlighted its multimodal capabilities this time, so for the second task, I selected a complex screenshot of a product prototype and asked it to perform visual recognition, understanding, and then iterative optimization design. This mainly tested two things: accuracy in multimodal understanding and competence in product design.

I gave Doubao 65 points. Its visual model's parsing was not detailed enough, with several obvious errors and persistent hallucinations. For example, it misread the growth rate of total customers as 3.6%, which was incorrect. The data for the Top 5 channels also didn't match—WeChat Official Account was marked as 26.6%, which was wrong. Additionally, its analysis stated that the '618 promotion event, which ended 15 days ago,' had actually just ended 5 days ago.

However, Doubao had one strength: its product design thinking was relatively clear. Since my requirement mentioned this was an enterprise AI marketing workstation, and the original product screenshot didn't prominently feature AI elements, Doubao actively emphasized them during iteration, enhancing user perception. This showed thoughtfulness.

I gave Codex 80 points. Its visual recognition was strong, with significantly higher accuracy and virtually no data errors. However, its product design was conservative, focusing only on local optimizations and lacking the novelty of Doubao's approach.

Later, I had them each output an improved version of the product prototype. This is Doubao's version, which clearly highlights AI elements.

This is Codex's version, with minimal differences and only local optimizations.

Score: Doubao Pro 65 vs. Codex 80

Scenario 3: Elevator Scheduling Game—Not Just 'Can It Move,' But 'Can It Move Well'

The third task was a coding challenge: creating an elevator scheduling game to transport as many passengers as possible within 120 seconds while minimizing waiting time. This task was comprehensive—it wasn't just about writing code but also involved visual presentation, interaction logic, scheduling algorithms, and software testability. Unlike coding games like Snake or Sudoku, whose source code is widely available online, this required some independent analysis and judgment from the model.

In the first direct output, I gave Doubao Professional 60 points. It implemented some basic functions—the elevator could move—but there were many issues. For example, passengers should have been on corresponding floors, but Doubao separated them, indicating logical flaws.

The status display in the passenger list on the right was too crude, lacking clear text states like 'Waiting / Assigned / In Elevator.' The class names for waiting and in-elevator styles were incorrect, leading to unstable status colors. The scoring animation existed but wasn't tied to specific floors or passengers, weakening feedback.

Another easily overlooked point: software testability. Doubao's debug mode was basically unusable. It also lacked refinement in details, such as residual overlays after pausing and resuming.

Codex's performance far surpassed Doubao, earning 90 points. Its scheduling logic was smooth, interactions were fluid, and the debug mode was functional. Deductions came mainly from minor details. For example, the elevator's boarding and alighting rhythm felt mechanical, and mobile adaptation wasn't as good as PC (single-column layout on narrow screens, button widths and font sizes weren't well-adjusted, requiring scrolling in core areas, affecting operational continuity).

Overall, Doubao could create a 'moving elevator' but lacked engineering capabilities in scheduling logic, state management, interaction feedback, and testability. Codex excelled in fundamentals, considering everything from algorithms to interactions to testability. Of course, this is only the first version of Doubao Professional—expecting it to match Codex, which has undergone countless iterations, is unrealistic.

Score: Doubao Pro 60 vs. Codex 90

Summary:

When is an AI tool worth paying for? I believe there are three criteria.

First, the hallucination rate must be low—this is the baseline. I fear most when it mixes true and false information, requiring me to verify everything myself, rendering it useless. For a tool you pay for, you expect trustworthy results, and Doubao still needs improvement here. If it doesn't meet a usable standard, I won't use it based on this alone. This issue must be resolved.

Second, it must have strong intent understanding capabilities. Previously, using AI required writing professional, complete prompts covering all aspects. Now, good models don't need complex prompts—they can understand your intent clearly from a sentence or two, which is a sign of a powerful model.

Third, it should have a high first-time success rate, minimizing the need for repeated iterations and modifications.

Additionally, I value the synergy of several complex capabilities, such as visual and video multimodality.

Based on these criteria, my impression of Doubao Professional after testing is: the direction is right, but its capabilities haven't fully caught up. At 500 RMB/month with a 60-65 point performance, it's definitely not enough.