Using Doubao for work? You'll have to sit at the kids' table.

葬AI·June 25, 2026

I held it together.

"I held it together"

ByteDance just held its big conference, announcing that Seed 2.1 has arrived — topping the charts again, another qualitative leap, another revolution. Now with an "office mode" that can actually get the boss's work done.

My first reaction: no way. Because Doubao has always had this reputation of "nice to use but not that reliable." You can chat with her, sing karaoke with her, ask her dumb questions from the "Dumb Questions Bar." But I've never seen anyone around me actually trust her with real work.

And now Sister Doubao is announcing a career change? No more playing around, she's a secretary now?

I ran some emergency tests right away to see what was going on.

After a full round of testing, I burned through nearly 200 RMB in API calls. The 69 RMB standard Pro version of Doubao hit its cap after just five tasks, forcing me to upgrade to the 200 RMB advanced tier.

The only time I've had spending this aggressive before was testing video models. Doubao, you've changed 😭

And this Doubao Pro looks identical to Trae Work? Which one are they actually pushing? Are we doing the Tencent horse-race thing now?

Test 1: World Cup Soccer Game

First up, coding. I hooked up Doubao-Seed-2.1-pro to Trae, ByteDance's own IDE — same source, same treatment.

After Fable 5 dropped, all these AI influencers on X were claiming they hand-coded games with Claude. Fable 5 isn't available anymore, but I figured it'd make a good benchmark.

So I asked Gemini 3.5 Flash, GPT-5.5, and Doubao-Seed-2.1-pro to each make me a "World Cup game featuring Team China."

Here's what GPT-5.5 came up with ⬇️

Damn, this actually looks like a budget FIFA game. You can sprint, tackle, shoot, even substitute players.

As for Gemini 3.5 Flash, here's what it made ⬇️

Now both teams are down to one player each, and all you can do is shoot — nothing else works.

Finally, here's what Doubao-Seed-2.1-pro gave me ⬇️

Kind of a pleasant surprise, because based on yesterday's tests from Xianyu, Seed's coding ability was basically written off as trash. Didn't expect it to actually produce a game.

But the downside: later on the ball just flew up and stuck to an invisible ceiling. Nobody could touch it. Forced draw.

Test 2: Feng Ge's Resume Website

Of course, making web games is a hype task. Let's see if it can handle actual work.

So I had Opus 4.8 and Doubao-Seed-2.1-pro both tackle a task: "Make an Apple-style website for blogger @Feng Ge Wang Ming Tian Ya, with 3D effects and parallax scrolling."

Opus 4.8 finished the page in 5 minutes ⬇️

Decent imitation of Apple's color palette and style. The parallax scrolling is there.

But the 3D is mainly just card tilting — barely noticeable, kind of lazy.

Then for some reason, with the exact same prompt, Doubao-Seed-2.0-pro took a full 20 minutes.

Here's the result:

Barely has anything to do with Apple's website style anymore, but damn are the effects flashy.

Seed approaches web design like a nouveau riche from the countryside — just pile on the bling.

The biggest problem: Feng Ge's photos, quotes, and works are all made up. Given that Feng Ge is a major creator on your own Douyin platform, you really couldn't find his info?

Or maybe this isn't Apple Feng Ge, this is Android Feng Ge.

Both the World Cup game and Feng Ge's website share the same flaw: they take foreeeeeever. These tasks took roughly four times as long as other models.

I suspect Seed is secretly "rerolling" for user satisfaction — running the task four times and picking the prettiest result to send me.

So Doubao-Seed-2.1-pro is kind of like pretty food — great for Instagram, rough as an actual meal. We're using AI to cut costs and boost efficiency, and now costs aren't low and efficiency isn't high. It's the most beautiful contrarian.

Test 3: North Korean Restaurant Vlog Editing

Next I tested visual understanding.

A while back I went to Goryeo Restaurant, that famous North Korean state-run spot in Shanghai. Shot a bunch of footage but never got around to editing it into a vlog.

So I dumped the clips on my computer and had Opus 4.8 and Doubao-Seed-2.1-pro each annotate, organize, and edit the footage, plus mosaic out the General's face.

First, here's what Opus 4.8 ended up with ⬇️

The subtitles are peak stereotypical "sis/bro" internet speak — doesn't sound like anything a human would actually say. But it did basically cover the General's face solidly.

Now for Doubao-Seed-2.1-pro — this round I used Doubao Pro's "Office Task" feature. It first organized my footage into these forms ⬇️

Looks very thorough and detailed, except a lot of it's fabricated.

For example, since the North Korean restaurant strictly prohibits filming during song-and-dance performances, there are zero videos of the performers. But Doubao claims I recorded multiple clips — basically framing me, trying to get me in trouble.

I also never ordered the Daidonggang beer or ginseng chicken soup she mentioned, never took photos with the North Korean staff. All Doubao's fantasies.

Truth is, a lot of people use Doubao for the emotional value, and that's fine.

But when you've selected "Office Task" and still pull the emotional value routine, that's false advertising.

Anyway. I then asked Doubao to edit these into a vlog and mosaic the General's face. Here's the result ⬇️

It does beat Opus 4.8 in some ways: the voiceover has more internet-native flavor; since Seed has its own voice model, there's pleasant narration; it even decides on its own whether to use original audio or voiceover based on the scene, giving it a nice rhythm.

But there's a fatal flaw: I told Doubao to mosaic the General's face. Doubao studied it for ages, and ended up nearly putting the mosaic on my face.

Left is Claude's, right is Doubao's

I'm deceased. Didn't they say the multimodal capabilities were god-tier? This is it?

ByteDance content moderators have been worrying about getting replaced by AI. I hereby announce: you can stop worrying, because AI can't even see the General's face.

I used to think ByteDance kept human moderators around out of humanitarian concern. Now it looks like the models just aren't there yet.

Humanity thanks Doubao.

Test 4: Ethics Variety Show Recap

Later I grabbed an episode of MILF Manor and had Doubao cut it into a Douyin-style movie recap short — another capability Doubao loudly announced as newly evolved.

They even had a promo case claiming it could digest two-plus hours of video in one go and output precise recaps. Let's see the real level?

Oh, MILF Manor is a famous dating reality show where 8 single moms and their 8 sons cross-date each other. I figured this complex relationship web would be the ultimate test of AI comprehension.

After multiple rounds of refinement, here's what Doubao-Seed-2.0-pro gave me ⬇️

At first glance, no obvious problems. But the narration and visuals are completely mismatched.

When it should show plot footage, it's all talking-head interviews — zero visual impact.

When the narration says "the eight sons are completely naked," everyone in the shot is fully dressed.

Though on second thought, maybe it's fine. Maybe Douyin users don't even look at the video, just listen to the audio. So maybe this feature really will put movie recap editors out of work.

Test 5: Mahjong Soul Auto-Play

Doubao's task mode claims it can control webpages and computers. Filling forms, buying tickets — small stuff I assume is no problem, no need to test.

So I had it open Mahjong Soul and play a round of Japanese mahjong for me.

Instead, Doubao played itself into an existential crisis. One moment confused about what tiles it held, next moment confused about what it just discarded. Turned mahjong into a suspense thriller.

Later I checked the replay. While Doubao was talking a big game to me, pretending to strategize, it was actually pure cosplay — discard whatever it drew, zero thought.

Why? Can't be deliberately scamming me, right?

After some digging, I realized: I thought Doubao was executing this mahjong task via real-time video stream input to observe and control the webpage. But it's actually using the same approach as the previous Doubao phone solution — high-frequency screenshots, then analyzing images to reason about the game.

This creates two problems:

The game state changes in real time. The board state in your last screenshot doesn't exist a second later. All that analysis is punching air.
Even fast analysis wouldn't save it, but as mentioned, Seed models are just too damn slow. By the time it finishes analyzing this round, 10 rounds have passed and someone else has already won.

Both problems exist in actual office scenarios too:

When Doubao is "helping" you work, it's responding to client requests from minutes ago, making click decisions that are already outdated.

Unless you have infinite time and a permanently static work interface, using Doubao to assist your work is self-torture.

Of course, the screenshot approach itself isn't the problem — real-time video streams definitely wouldn't cover costs and would hit privacy and compliance red lines. At the end of the day, Seed 2.1 is just too slow.

Test 6: Modeling the Commander-in-Chief

Yesterday's marketing also featured a seemingly impressive 3D modeling case, so I tested that last.

I connected both Opus 4.8 and Doubao-Seed-2.1-pro to Blender via MCP protocol, requesting a Pop Mart-style Trump figurine, highest precision possible.

Results below ⬇️

Left is Opus 4.8's, right is Doubao-Seed-2.1-pro's. Honestly I can't really judge this one — vote in the comments, fam.

I also had Doubao-Seed-2.1-pro generate a gray mold version for 3D printing, and had Xianyu print it out ⬇️

Without color, it's kind of hard to tell this is a person.

And according to Xianyu, Trump's MAGA hat is all jagged edges, both feet are fused together — has that post-nuclear-wasteland aesthetic.

Didn't see any extraordinary modeling capability. Someone want this? I'll raffle it in the comments.

Overall, the Seed 2.1 update isn't perfect. It's squeezed its way into the "office"赛道.

But ByteDance's internal AI business is highly integrated, so the Doubao app and Seed model aren't a pure wrapper relationship.

For example, Seed's video analysis isn't that strong, so the Doubao app has no video upload button at all.

And since Doubao users love video-calling and singing with AI, Seed developed the awkwardly-named full-duplex tech Seeduplex — just so users won't interrupt each other while chatting with Sister Doubao.

This was dimension-reducing in the first half of the AI wars. No matter how much competitors improved their GPQA Diamond or HumanEval scores, it didn't matter — regular users can't feel that 0.5-point difference. But Doubao's comfort and convenience is tangible.

Doubao got the application layer right first, capturing user mindshare directly.

Users even actively accommodate Doubao's shortcomings. All over Douyin: Sister Doubao personifications, "I'm telling you in the most straightforward, direct, no-bullshit way that I held it together," and the always-apologizing "Doubao personality type."

At the core, this is all because the Seed model isn't strong or accurate enough — but Doubao users have reframed it as a kink, as cute.

ByteDance hasn't done anything about it. I figure half the time they have no choice, half the time the damage isn't that serious right now.

But now the AI wars have entered the second half, and everyone's rethinking applications. This round isn't about companionship, search, or ordering milk tea — it's pure productivity, pure workplace grind. Even Tencent stopped using Yuanbao to pull people into group chats, they're working on something called WorkBuddy now. You feel me?

This is high ground Doubao hasn't occupied yet. Everyone's back to equal footing.

Regardless of how competitors are actually performing, regular users are starting to wonder whether domestic AI can help them work. Once they discover Seed can't actually handle office tasks, they'll get suspicious. Sister Doubao has to throw on a suit, pretend to be a white-collar worker, and respond ASAP.

This Seed update is basically paving the way for this new "office"主旋律. The new capabilities in 2.1 pro are all productivity-oriented. In Doubao itself, they literally labeled 2.1 pro with "Office Task." It's practically whispering in users' ears: use me to save money, use me to make money.

But can you really?

There are plenty of issues. When I used Trae directly with Doubao-Seed-2.1-pro, the output was decent, but one video took a full day to edit.

Switch to Doubao task mode, and the work got done fast — but the output was barely usable.

So we're doing the can't-have-both-fish-and-bear-paws thing here.

Seed's internal tension between quality and speed remains unresolved. Far from a hexagonal warrior. I suspect the Seed team knows it's not perfect either, but got forced out by the Doubao Pro launch timeline.

So my takeaway is: Sister Doubao has entered the workplace, but her role is programmer morale officer — mainly providing emotional support.

When Doubao was a toy, "read and replied nonsense," "Doubao personality type," "only provides emotional value" — none of this mattered, it even added human flavor to the product.

But if Doubao wants to become a real office assistant, all these cute quirks become red flags that scare users away.

The worst-case scenario: users still use Doubao, just not for work. Then nobody pays, Doubao burns cash without revenue — that's the sky falling.

Doubao, you used to just need to listen to your master. Now you really gotta work hard to complete your master's tasks.

Otherwise Seed might end up like that mahjong game it played for me — unable to win at the table 😭