The August 2026 AI Build List: What to Actually Use for Coding, Images, and Video
Recently I've done a few livestreams, and the most common questions in the livestream room and the public account backend are about which AI to use for programming, image generation, and video, and which AI is the strongest.
Actually, the past month has truly been a month of major AI updates. Programming, images, video—it's no exaggeration to say each track has undergone a complete overhaul.
So today, I'll just give everyone an updated version, with conclusions by track. Pick according to your own use case.
You can think of this as an August 2026 edition AI build list.
1. A Few Clarifications First
First, I'm not giving conclusions from a self-media perspective, just scraping some info online.
I'm a frontline AI developer, spending nearly ten thousand a month on various AI subscriptions. Although there may be some personal preference, I'll try to be as objective as possible, just for your reference.
Secondly, Anthropic's Claude Code no longer has the cliff-like lead it once did. AI capabilities are gradually being caught up. In fact, if you've used it heavily these past few months, you'll find that while Fable 5 is indeed good, the reception for Opus 4.7, 4.8, and even the current Opus 5 has been very poor. Sometimes you get the illusion that this company's previous lead was simply because it chose the right track and enjoyed the dividends, but its relative rate of progress since then is truly unimpressive compared to others.
Coupled with Anthropic's corporate values and level of arrogance, I now sincerely suggest that Claude Code is no longer a must-have. Although it's still very strong, its status is incomparable to three months ago.
If you're still using Claude Code, no problem. But if you haven't started yet, you can just ignore it now; there's no need to force yourself to use it.
And as I said in a previous article, future AI competition may be more about the ecosystem, and Anthropic is weak in this area.
2. AI Programming Recommendation: Codex + GPT-5.6 Sol
If you want to use AI to build products, tools, etc.—that's the AI programming scenario—then without a doubt, Codex + GPT-5.6 Sol is the king of programming. Moreover, OpenAI has been frequently resetting Codex recently. In the past month, the Codex subscription plan has likely been the optimal choice in terms of both capability and cost-effectiveness.
But this is also OpenAI's marketing strategy, mainly to grab users. As the user base grows, resets definitely won't be as frequent as before, but the capability level is genuinely formidable.
Plus, there's OpenAI's ecosystem.
Codex is basically in the top tier of current Agents.
Its own GPT Image 2 is still the strongest image generation model to date, and image generation shares quota with Codex.
In voice, whether the previously open-sourced Whisper or the current GPT-Live, it's still the strongest.
Not to mention video; although Sora was shut down, when it first launched, it basically crushed all global video models.
In other words, OpenAI's AI ecosystem is very comprehensive, and its technical capabilities are very strong.
This means a single Agent, Codex, can easily handle any scenario.
3. Best Domestic AI Programming: Kimi Code + K3
When K3 first came out, I wrote a separate review article, because when domestic models first emerge, they're often accompanied by various hype articles claiming they crush everything. So everyone is used to it and wonders if it's just another exaggerated marketing push.
After using it myself, my review reflects my genuine experience.
I'm on Kimi's highest-tier plan. After deep use of K3, I can tell you that Kimi Code + K3 is far stronger than everyone imagines. People in China have underestimated K3's capabilities.
Now, Kimi K3 and Codex are both my main development environments, because one subscription plan isn't enough. Not to mention frontend capabilities, even in some complex programming tasks, actual tests show K3 is very competitive.
So, if for various reasons you can't use Codex, I strongly recommend Kimi Code + K3.
Of course, Kimi's problem is insufficient computing power; many people can't grab a subscription plan, plus the quota isn't enough for heavy use, and the speed is relatively slow.
4. Top-Tier Image Generation: GPT Image 2
In the image space, GPT Image 2 was released on April 21st, and it's been almost four months now. Its number one spot on the text-to-image arena leaderboard hasn't budged.
Google's Nano Banana is also good, but the Gemini ecosystem is too weak now; it's basically the American version of Doubao.
However, Google's problem is its own team issues. A few days ago, there was a major internal shakeup; the leaders in the AI technology field left as a group to start a company. Google made very significant organizational restructuring adjustments. This isn't a bad thing; give it some time. But for now, it's not a recommended choice.
Musk's Grok just released Imagine Image 2.0. According to official data, it ranks second globally on both the text-to-image and image editing leaderboards, and it's cheap.
Previously, Grok was never favored and had very few users. But since acquiring Cursor and merging into SpaceXAI, its recent AI progress across the board has been rapid. Recently, whether in AI programming models, images, or the video ecosystem, there has been progress.
The biggest advantage is that Musk doesn't lack money or computing power. If you've used Grok, your first impression is that Grok's throughput speed is damn fast. The output speed is like a machine gun, visibly fast.
So the Grok family ecosystem is worth watching; it has great potential later on, but it's not the first choice right now.
Of course, in the image field, domestic players like Hunyuan and Qwen have also released their latest updates, but currently, they still can't compare with GPT Image 2. Again, if it's not necessary to switch everything just for image generation alone, the Codex ecosystem is currently the optimal choice.
5. Video Duopoly: Seedance 2.5 and MiniMax H3
Seedance 2.5 excels in finished video capability. It generates 30 seconds of native video in a single pass, no stitching or super-resolution needed, and supports multi-round extension, theoretically able to continue to several minutes. Don't underestimate this 30 seconds; the industry average is still 5 to 15 seconds. This pushes AI video from "generating素材 (raw material)" to "directly outputting finished clips." You can feed up to 50 reference assets at once—30 images, 10 video clips, 10 audio clips mixed freely. It also supports precise timestamp-based editing, allowing you to change any specific second.
Currently, ByteDance's Seedance 2.5 is undoubtedly the king of the video domain, but the downside is that it's expensive.
MiniMax H3 excels in being open-source and cheap. The weights were directly open-sourced on August 3rd. 2K resolution, 15 seconds, with native stereo sound; picture and sound are generated together, no post-production dubbing needed. The API price is $0.13 per second for 2K, which the official line says is one-third the cost of mainstream models. If you have your own GPU, ComfyUI supported it the same day; pull it down and run it locally, and the marginal cost is just electricity.
My actual tests show H3 is also very strong, but there's still a gap compared to Seedance 2.5. However, H3 is open-source, which means very soon, the video model capabilities across the entire field have a chance to close the gap, and it might drive down the price of video generation.
We should be grateful to every open-source AI company; they deserve respect.
6. Just Copy This: Build List by User Persona
Finally, conclusions by user persona. Find your match.
If you're just doing daily office work, domestic users should go with Workbuddy without thinking. It has a bunch of built-in skill toolkits for workplace professionals, ready to use out of the box, with a low barrier to entry.
Developers who make a living from code, if conditions permit, just go with Codex + GPT-5.6 Sol.
Domestic developers on a budget, Kimi Code + K3. If you can't grab a subscription plan, check out some domestic cloud service providers.
For self-media creators, designers, and operations staff, GPT Image 2 is the main image tool. If volume is high and cost is a concern, check out the latest Grok Imagine Image 2.0.
Short video creators, Seedance 2.5. The 30-second direct output is currently in a league of its own, but the cost is relatively high.
Tech teams needing batch video output or with hardware resources wanting to self-deploy, MiniMax H3.
One tip to add about video generation: video production is actually divided into Motion Graphics and creative asset-type videos. The former is code-engineered video production, giving you pixel-perfect control; the latter requires video large models for generation, which is costly and involves a lottery-like element.
If you have some technical ability, you can combine Motion Graphics engineering capabilities with the creative aspects of video large models. Done well, this can drastically reduce video production costs while maintaining equivalent video quality.
Recently, our team has spent a lot of time researching this area, and we've now updated the AI video tutorials on our planet. If there's a chance later, we can share real-world video quality and costs.
Finally, to be honest, there's no "graduation build" in the AI industry, only the most suitable configuration for the moment.
The above build list is just my current recommendation, for your reference.