Transcription
Recently, too many new neural networks have been released. Leaderboard records were broken almost every week. Google finally blew away all metrics with its Gemini 3 Pro and the latest Flash. The new Opus returned Anthropic to leadership positions in coding, and OpenAI systematically increased power with the release of GPT 5.1 and 5.2. In this video, I will compare the top models from the leaderboards on VIP coding tests, making them complete complex projects in one prompt, and try to find out who is currently the best in programming. Let's go! Hello, my name is Oleg. In my everyday life, I am a developer, and on this channel, I am a vibe coder. Here I test neural networks and tools, build micro-projects and automations. I see what really works and what is just hype. Which models will we compare today? From Anthropic, I will take Sonnet 4.5 and Opus 4.5, as well as a similar pair of models from Google – Gemini 3 Pro and Flash, plus the flagship from OpenAI, GPT 5.2 with the thinking mode turned off to the maximum, and turned on to extra high. Also, for interest, I will add the fast Composer 1 from Cursor, two powerful open-source models Kimi K2 and GLM 4.7, and the Chinese titan Qwen 3 Max. I am not testing the over-downloaded GPT Pro from OpenAI, as it is not included in any coding tool subscription, even for two hundred dollars. Accordingly, it can only be accessed via an API key, and via API, this model consumes tens and hundreds of dollars in minutes. I tried to test it, but with minimal progress on the task, it consumed about fifteen bucks. At this point, I decided to skip its testing. Unfortunately, my budget cannot yet afford vibe coding on this model, and it is also really very slow for coding. As a development environment, I will use Cursor today, as six out of nine models are available in it out of the box, so their conditions will be as identical as possible. For the rest, I will use Cline with API key integration, as they do not work in Cursor at all. And as a provider for all text models via API today, I will use Polza AI. This is an excellent API aggregator that allows you to test and use hundreds of LLMs, easily integrating them into any project. If you are building your own AI services, experimenting with LLMs, or building automations in the style of HeyGen, then I recommend taking a closer look. Everything works without VPN and foreign cards. The API interface is compatible with OpenAI, so there are no problems integrating Polza AI into any project. For example, in Cline, I simply select the OpenAI Compatible provider type, enter the base URL, API key, and the desired model. Done! And the integration works. Similarly, everything is done in code if you need to implement some AI service. In general, if you have been stopped by complex payment methods, VPNs, fear of bans, and other problems, then try Polza AI. The service really simplifies life. Plus, they have also added image generators. You can use NanaBanana, GPT Image, Seedream. This is very useful for many AI projects. The link to the service is in the description. Register, test, integrate into your projects. And now, let's move on to the first test. To warm up, let's start with a simple task. Lately, I've started to feel that many things in life are done too complicatedly. You need to make multi-step processes, observe timings, get certificates and documents. All this is quite difficult and unpleasant to keep in mind. However, if you imagine that you are not just getting a strange certificate for a visa, but completing a website quest from an RPG fantasy game, then anxiety disappears. You are like just a player completing a task. And on this topic, I came up with the idea to create a basic to-do list, but in a fantasy style. It should have achievements for completed tasks, items, and the atmosphere of Middle-earth. I think this project is perfect for the first warm-up test of neural networks. For the vibe reference, I selected these pictures on Pinterest, so that all models could understand the plot, colors, and style of the pictures, I asked GPT to describe each of these images in great detail. I got this prompt, supplemented it with my own functional wishes, and at the end added a threatening text with the conditions of today's tests. And the conditions are as follows: for each task, I give one starting prompt with a description of the entire project. Then I check the implementation result, and if something is wrong, I make up to five additional calls to the model with a detailed description of the found bugs. After the fifth correction, I consider the result final. Describing these rules in the prompt makes the model test the result independently and stop only when it is truly confident that everything will work correctly. This feature is useful to save time on manual task refinement where the probability of error is very high. Here is the prompt for the fantasy to-do list. What's important is that I want the quest icons to be made using emojis generated by the LLM based on the task text. Plus, achievements for completing quests should also be given with text from the neural network. And for models from Cline, I added a note where to find the rules that work by default in Cursor. In these rules, I described working with MCP Context 7, plus how to properly manage Python code if it exists. The full text of the prompts, projects, and rules will be in the video description. After launching the prompt for each model, I got an interesting result. Let's go and look from the cheapest to the most expensive models. Composer created a scary design. The program had no achievements or items. GLM 4.7 created a working product. Quests are created, but not marked as completed. No achievements. Kimi K2 creates tasks, but does not complete them. And I want to note that Kimi K2 kept freezing after every action for some reason. This was incredibly annoying. Gemini Flash works more or less, although there are no items. Qwen 3 Max could not finish the code, getting stuck in the middle. In the GPT project, everything works, the design is normal, there are mechanics. Gemini Pro – strangely, but creating quests after all the fixes still didn't work. Sonnet works fine, the design is unknown, but overall it's okay. And Opus did everything as it should. Achievements, items, icons, cool design, fonts – great. Here are the ratings I give the models at this stage, and let's move on. In the second task, I decided to create a system for collecting information from Reddit. Reddit is something like a forum with a wide audience worldwide. People there often write about their problems, comment and discuss pain points, share solutions. Reddit is a good source of ideas for startups. Based on user desires and problems, you can come up with many good products. I decided to create a system that collects the texts of posts and comments to them, then processes them through an LLM, identifying the author's problems, calculating embeddings based on the problem text, identifying the vector of meaning, and building a map of found problems based on these vectors to visually find clusters of similar pain points. I decided to collect data by simulating browsers with Playwright just in case, plus using proxies to avoid bans. Without proxies, such systems always hit one IP, which is very easy to block on the source side, and we need to avoid this. Proxies will allow us to use many different IPs in the location we need. However, the problem is that many proxy providers have dirty proxies. This leads to a lot of false blocks in the form of captchas, redirects. In such cases, it seems like a bug in the code, but in reality, it's in the infrastructure... Therefore, for such tasks, I use NodMaven. They claim ninety-five percent clean IPs, sessions can last up to twenty-four hours, plus there is normal filtering by proxy quality, location, and other parameters. Very convenient when you need something specific. For my project, I will take the USA, California, a random city. I think Reddit will like my location, as it is primarily oriented towards Western countries. And another cool thing: NodMaven has a speed filter that allows you to open sites faster, almost at the speed of home internet. This helps in cases where time is important. If you also want to use high-quality proxies, then follow the link in the description. With my promo codes, you get a fifty percent discount for the first fifty users and a hundred percent bonus traffic. Here is the project prompt we got. Let's look at the result. Composer did not cope, after five fixes it does not work. GLM 4.7 starts, mines something, but ultimately gets nothing. Kimi K2's frontend opens, but nothing starts with the "Start" button. Gemini Flash fixed everything after two prompts. The map is a bit inconvenient, but it works. It contains both comments and posts. Qwen 3 Max did not start after fixes, the system does not find information. GPT's system worked right away. After two prompts, minor details like zoom and dot size were fixed. Gemini 3 Pro launched, but nothing was found. Hmm. Sonnet works fine, but comments were not loaded into the system. And Opus did everything after three additional prompts. Here are the ratings for this test. Let's move on! Lately, I've been using Dokploy a lot for managing projects on servers. It allows you to quite conveniently deploy any images in a couple of clicks by connecting your GitHub account, however, I haven't found a way in it to deploy public repositories without logging into an account. Sometimes I need this when setting up other people's servers, as I don't want to give access to my account, and Dokploy also lacks alerts if something goes down, and any built-in editor to deploy simple custom images. Yes, you can try switching to Coolify, it seems to have a wider set of features, but let's try to develop our own custom admin panel for my request in this test. Here is the prompt I composed. I want to receive alerts in Telegram to quickly respond to failures, and everything else I listed earlier: GitHub, editor, plus domain and HTTPS setup. But I won't check domains yet, as it's quite difficult to do on a personal computer. You need to put it on a server and test it there. Launching coding agents. Composer created a project where nothing really works, nothing is deployed. GLM 4.7 launches something, but there's nothing in the browser. Kimi K2 could not get anything working after all the fixes. Gemini Flash has a GitHub connection, but private repositories are cloned empty. Build logs are not displayed much, but public projects launch if the name is unique. The design is ultra-default. Qwen 3 Max – the design is terrible, a super simple project creation form, and it crashes with an error. I fixed this for an hour, consuming about ten bucks, but didn't fix anything. GPT – almost everything worked right away. After two prompts, everything started working perfectly. It finished faster than everyone else. Quite a pleasant design, understandable interface. Overall, cool. Gemini 3 Pro. After all the fixes, everything more or less started working, only if Docker doesn't build somewhere, it crashes, nothing is visible in the interface or logs. Plus, the design is ultra-bad, very off-putting. Sonnet – after five fixes, everything works: GitHub integration, file editor, start, stop. Only alerts didn't manage to get working. And Opus – after five fix prompts, it works perfectly. All types of project launches function: public, private repository, code editor, log search with and, or, and parentheses, start, stop, Telegram alerts. Here are the scores for this test. Let's move on! And now a really useful test. I often think about translating my videos into another language, for example, English. With modern technologies, this can be done automatically, however, almost nothing of the available solutions works well enough. HeyGen has automatic dubbing, but customization is very weak. Other services are even worse. Of the interesting ones, I only found Dubly. For a hundred euros, you can translate a thirty-minute video with a more or less decent editor, but they also ask for a hundred euros for lip-sync, so that the mouth on the video moves in sync with the translated voice. For me, this is a bit too expensive. Plus, lately, I have studied enough models, practiced creating pipelines, so I think I can build my own pipeline for video dubbing into any language, plus an editor for correcting translation and voiceover problems. The system will work like this: the audio will be extracted from the source video, then it will be transcribed using ElevenLabs, translated with an LLM, and voiced using voice cloning in Minimax, and lip-sync will be done in Pixverse. Fortunately, I still have code examples from the short video generator where I integrate with Minimax and Pixverse. I will add them to the initial folder so that the models immediately understand how to work with these APIs. Here is the prompt that came out. In it, I ask to create an editor with step-by-step generation. In it, you can edit the original text, translations, regenerate the voice, and see what happens. Let's launch! Composer produced what seemed like a working project, but in fact, the transcription doesn't return any words, and everything else is meaningless. GLM 4.7's system works, there is a result, but the preview at all stages is not displayed. Vibe Coding - is a new way of creating programs, where you forget that code even exists. You just describe what you want in plain language, and the neural network writes all the code for you. Today I am implementing a project for automatic video cutting based on a script, and I will do it with an absolute vibe coding approach, without looking at the program files at all. Kimi K2 could not even start transcription after all the fixes. Gemini Flash – after five fixes, transcription and translation work, then it crashes on voice dubbing. Qwen 3 Max could not solve the transcription. GPT – after two fixes, the result is excellent, everything works clearly and predictably. There is progress in percentages for each task. The voice is a bit strange, but probably due to the wrong Minimax model, it can be corrected later. -Vibe Coding is a new way of creating programs, where you forget that code even exists.... You describe what you want in plain language, and the AI writes all the code for you. Today I'm implementing a project for automatic video cutting based on a script. Gemini 3 Pro. After fixes, the system generally started working, but without selecting an existing voice ID and a proper preview of the final result in the browser. After downloading, everything is okay, it can be used. Vibe coding is a new way of creating programs, where you forget that code even exists. You simply describe what you want in plain language, and the neural network writes all the code for you. Today I will implement a project for automatic video -cutting based on a script text. -Sonnet – after five fixes, it still couldn't handle the "transcription failed" error, and Opus. The program worked immediately without fixes. There were only a couple of minor problems. After two fixes, everything works great. Vibe coding is a new way to build software, where you forget that code even exists. You just describe what you want in plain English, and the AI writes all the code for you. Today I'm building a project that auto cuts video to match a script. Here are the scores for the current test. And let's move on to the traditional final LLM model test. We will test the spatial reasoning of text models by making them create 3D models in Blender using MCP. If anything, I showed how to connect such MCP in this video. In previous Blender tests, I always asked to create a Soviet panel building, but today we have very smart models. I want to create something a bit more complex, so I propose to create a 3D model of the Japanese Matsumoto Castle. It's beautiful, not super complex, but not entirely square either. Let's go, let's implement it! Here is the prompt that came out. Let's launch, you can see the result yourself. I like the results of GPT and Opus the most. And here are the final scores. GPT won absolutely all tests. It seems to be precisely due to its extra high thinking mode. It thought for a very long time. Lately, there seems to be an obvious connection between the model's thinking time and the quality of the final result. Opus, in this regard, apparently doesn't strain itself as much, although it still produces excellent results. The difference between these models in today's tests was practically insignificant, purely in details and micro-moments. The fastest of all models was definitely Composer, however, its results are not super cool. I think it's cool to use for small, specific tasks where you don't need to rework an entire system. Gemini three Flash also surprised me. Its cost and results are in a very good balance. This seems to be the best option for budget AI coding. And I want to note that in such tasks, the combination of tool and model is very important. Kimi K2, Qwen, and GLM, for example, connected to Cursor and even responded to some prompts, but for some reason could not work with the tasks of today's tests and could only launch normally in Cline. Perhaps, by analogy, models from Anthropic will work more effectively in Claude Code, and Qwen in some Qwen Code, possibly, but this is a separate topic for other tests. And what else I noticed today is that you can implement small MVP startups for yourself in one prompt, if there is normally described system logic and examples of using the necessary APIs. Now the bottleneck of development is not code or prompts, but studying documentation, testing, thinking through algorithms and architecture. These stages take the most time. Cool! Now I want to try to create real products with UI and monetization in one prompt. It seems that this is already quite possible now. In general, here are the conclusions and results. Such a start to 2026. It's interesting what will happen next this year? Will we see any radically new models, maybe new architectures or approaches? We'll see. In the meantime, subscribe to my channel. Here I will continue to test models, compare tools, show my small projects. Also, subscribe to my Telegram channel. There I sometimes write useful things about AI and coding. And write to me in DMs if you need help with automation or implementing neural networks. Well, that's all, see you!