Maximizing Efficiency from AI LLMs: Which Model is Best for What?

Comparing frontier LLMs, benchmark evaluations, and optimal model selection for marketing, engineering, and SEO tasks.

Speakers: Zafer Yıldız

Content author

Author
İren SaltalıinZeo Talks & AI Talks

09 Jul 2024

Video Transcript

⏱ 18 min read• 3,168 words

Read the full speech or switch languages

Hey everyone, welcome to Reo Talks. Today, we’re diving into AI. I’ll show you how I use AI in my daily life. I’ll share how I integrate it into my work. I’ll cover the tools and models I use most. I’ve planned a session with use cases to inspire you. I work on the Zeo Analytics team. We’ve covered analytics in past Zeo Talks. Today, we’re shifting focus from analytics to AI. Let’s get started. First, LLMs are large language models. They’re AI we can chat with. ChatGPT is the most famous one. We use chat interfaces. But these tools have many different features and use cases.

Sticking to one AI model isn't sustainable. This world moves fast. Something new drops almost every day. The models you see here might change by the time you watch this. This is the current landscape as of March 28, 2025. I adapted this table from Ozan on the Zeo team. I translated it and integrated it into Google Sheets. We’re looking at the top models here. We’ll cover their strengths and best use cases. I’ll show you which models to use for specific tasks. A great coder isn't always a great thinker. Some models excel at reasoning over raw coding. Let me introduce you to a few key players.

GPT-4O is your daily assistant for task automation. It’s best for general tasks, planning, and reminders. Use it for daily productivity and routine workflows. It even handles scheduled updates and automated reminders. OpenAI’s "O" series models are built for deep reasoning. They aim to provide much smarter, thought-out responses. This includes O1 Mini, O1 Pro, and O1 Mini High. O1 Mini is faster and more cost-effective. It’s perfect for quick queries via API. It excels at coding, programming, and complex math. Use it when resources are limited or speed is critical.

O1 Mini is great if you need quality without extreme detail. It’s 3-5 times faster than O1 Preview. It’s also 80% cheaper. Knowing these details is the whole point. Don't get overwhelmed by the sheer number of models. Most of these require a paid subscription. Knowing these features helps you optimize costs. It also ensures you get much higher quality results. I pick my model based on the task. Coding needs a different model. Searching for a recipe needs another. Don't just stick to one model. Avoid rote usage.

Things move fast, I know. OpenAI drops a new model and habits change overnight. But remember the basics. It’s all just Q&A. Good prompts get good answers. Or you can build advanced automations via API. Just keep the costs in mind. Know what you're using. DeepSeek R1 is a new Chinese AI model. It’s actually really good. You get fast responses. It’s solid for structured problem solving and coding. Gemini and Claude 3.7 Sonnet have passed it. Still, it’s a strong contender. It breaks down problems well. It handles relationship analysis nicely. Gemini is my go-to for speed.

Except for the latest Gemini Pro. We’ll see that later. Gemini is fast and has a huge context window. Think of the context window as the model's data capacity. High context windows let models analyze chat history or massive data sets. Token context windows are key. I'll explain exactly what that means in a second. Next is Grok. It’s one of my personal favorites. I use it on my phone and in my browser. The free version has very few limits. You can use it a ton without paying a dime.

Honestly, it gets the job done. Unless you're obsessed with AI like we are, Grok is a solid choice. It’s great for deep knowledge, logic, math, and coding. Use it for complex problems that need deep thinking. Like Gemini, it has a 1 million token context window. Grok is definitely in my top three models right now. I highly recommend trying Grok. It’s also great at image generation. You’ve probably seen those AI-generated political images all over X lately. It creates images of anyone without restrictions. Maybe it’s Elon Musk’s free-spirited approach. Grok really doesn't have many red lines in that department.

It generates visuals more freely. ChatGPT limits you there. It doesn't mind depicting real people. It doesn't mind changing their look. Grok doesn't have those limits. It's a controversial topic, honestly. Perplexity is more of a tool than the models I just mentioned. It replaces Google for AI-powered search. We’ll see an example soon. It raised serious funding. It has a very bright future. Claude 3.7 Sonnet is Anthropic's model. It’s neck-and-neck with Gemini 2.5 Pro in coding benchmarks. For coding, I go straight to 3.7 or Gemini. They are my go-to models right now.

It has hybrid reasoning and perfect coding skills. 3.7 writes code beautifully. It handles web development and complex logic well. It has great extra features. NotebookLM is a tool just like Perplexity. It’s a Google tool. If you’re a student or an academic, it’s perfect. Even if you just want to learn something new, it works. You can use a video or any source. NotebookLM is the right choice for you. It processes huge texts or video content easily. It gives you a clear output. We’ll see that in a moment. Finally, there’s GPT-4.5, another new model. It has an improved knowledge base. It features better conversation skills and fewer hallucinations.

That’s a feature of GPT-4.5. Honestly, it’s not my favorite model. I just can’t get into it. There are benchmark sites that rank these models. LiveBench AI is one of them. This site or LM Arena AI works well. It’s not loading in my browser right now. Maybe it’s a temporary glitch. I use this one a lot. It ranks the models I just mentioned. You can see live scores after they run specific tests. Need help with coding? Data analytics? A math problem? These sites tell you which model to use. For example, if I have a coding task, Gemini 2.5 Pro is currently the best.

I mentioned Sonnet earlier. I usually get better results from Sonnet. But apparently, the new Gemini 2.5 Pro is crushing it for code. It’s also leading the global rankings right now. Like I said, you have to keep track of these updates. You need to see which models are performing better. This list is only accurate for today. The rankings will change in a month. Let’s talk about tokens. You need to watch these when using LLMs. Tokens are basically the numbers generated in the background. Every interaction has a specific token count.

Every new message increases the token count. The assistant's replies also add to it. The model's response directly impacts your token limit. Even your source files consume tokens. This is the right tool for that. Let's move on to practical use. There are so many tools here. You have direct access to LLMs. You can go straight to ChatGPT, Claude, or Gemini. But tools like ThinkBuddy offer multiple models. You pay one fee for access to all of them. ThinkBuddy is a solid choice. Poe is great too. You can pick any model you want. Let's test Claude for coding. I already prepped a prompt.

I asked for a Python function to reverse a string. ThinkBuddy generated the code and an example. Claude 3.7 Sonnet provided this answer. I could have used Gemini 2.5 Pro for the same task. Now, let's look at Poe. I used Poe heavily for a while. I really like its features. Just like ThinkBuddy, you get tons of LLM models in one place. You can also build your own bots here. There are many models, apps, search bots, and image generators. You can use all of them. Just like in ThinkBuddy, you ask a question and get answers from any model. It’s that simple. My favorite feature is the bot builder.

When you click create, you see the types of bots you can build. You can create a bot to answer specific questions. No need to rewrite your prompts every time. It’s a lot like ChatGPT custom GPTs or AI agents. You just build bots. I already started drafting a prompt for one. You just enter your prompt here. I tell it to take complex topics like black holes or AI. Then it explains them simply for a 12-year-old. You set it up and give it a name.

I call it "Explain it simply." You add the prompt and a name. You can even embed specific sources to train it further on a topic. You can share these, but I usually keep them private. It offers suggestions or lets you share with others. I use it differently. I build custom bots for specific sectors. I can share them with friends easily. You just build a bot and publish it. It asked for a name. I'll hit publish. Let's ask: what is photosynthesis? I want simple explanations. It could be quantum physics or AI. I asked about photosynthesis. It explains it like I'm 12.

Think of it like cooking breakfast. It uses simple analogies. You can use this in Poe for anything. Ask about quantum physics or AI. It will give you the same simple answer. It breaks things down perfectly. It’s a great learning technique. Now, let’s move to TypingMind. You use AI models here via an API. You don't get all LLMs built-in. You must add your own API keys. Once added, you can use every model. TypingMind also has great plugins and agent features. This is one of my go-to tools. We won't get answers yet. We need an API key. Let's test it. I have a simple question here. I'll copy and paste it. It says I need an API key to get answers.

I generated a Claude API key for you. Once I hit save, I can query this model via Claude. I'll hit send. There we go. It's working now. I asked what compound interest is. It gave me a great summary. You can add three main API keys here. You can use Gemini, Anthropic, or ChatGPT. For other models like DeepSeek R1, you need a custom setup. It isn't pre-loaded here. There are agents too. They work just like the bots in Poe. But there's even more to it.

You can add plugins here. You can assign a model or train it just like in Poe. You can create a persistent assistant by entering your prompt. You can save your frequent prompts in the prompts section. Think of it as a prompt notebook. Browse prompts shows you the built-in TypingMind prompts. Need human-like content? Use these prompts to write emails or copy fast. There are tons of models here to choose from. You can use Google, Anthropic, and OpenAI here. TypingMind has so many features. I’m still discovering them. ChatGPT is just the standard chat interface. Ask your question and get an answer.

ChatGPT has two main search functions. It can do web searches or deep research. You can switch between different models. I have a prompt for deep research. Ask how a microwave heats food. It gives you the answer instantly. It doesn't need to think hard. It’s a simple question. It’s already in the LLM's database. It doesn't need live data for this. Things like photosynthesis or inflation are easy. You get fast answers for those. But for more complex stuff, try this prompt. I want a detailed weather report. It searches the web for me. It scans the web and gives me the weekend forecast. It even cites the source it used.

Select deep research or web search options based on your prompt. This gets you higher quality answers. Let's use deep research. What are the pros and cons of EVs as of March 2025? I have a prompt about charging infrastructure. It asks me to clarify things. It wants to know the brand, model, or language. I just say yes. One yes. Want a comparison? Yes. Three yeses. It starts a comprehensive search. This might take a while. It asks for details I missed in my prompt. Then it starts a deep dive.

It will provide a detailed output. Let it keep researching. Let's head to Gemini. Gemini is great for coding right now. I can paste the same prompt I used in Claude here. It's thinking. This is a simple code. Use Gemini for complex projects or app development. By the way, ChatGPT has checked two sources so far. This can take a while. It's browsing live and reading sources. It analyzes the data to give you a result. Don't search manually. Just ask an LLM to do it for you.

That workflow is pretty outdated now. Googling, clicking results, and analyzing data is old news. LLMs do all that for you today. Let’s look at Grok. It’s one of my favorite LLMs lately. There are two modes here. DeepSearch and Think. Or you can skip both for speed. Like that microwave question earlier? It doesn't need deep research or heavy thinking. Just ask and get your answer. Need some strategy? Use the Think mode. Here is my prompt. Give me 10 creative ideas for a small bookstore to get more customers. I’ve enabled Think and I’m sending it now.

It starts thinking. You can actually watch the thought process. It’s basically consulting with itself. You’re giving it time to process the prompt. Let it keep thinking for a bit. Back to ChatGPT. It’s still researching. It’s pulling detailed info on electric cars. I can see the sources here. It’s up to 14 sources now. It’s doing the heavy lifting and showing me the details. Back to Grok. It finished thinking. It took 32 seconds. Let’s see what it came up with. It listed 10 creative ideas for a small bookstore. Looks good.

Book clubs, author meetups, signings, and story hours. Workshops, loyalty programs, and book recs. Coffee pairings and social media contests. It all makes perfect sense. It gave me 10 detailed suggestions. I’m actually pretty happy with this. Now, I’ll try DeepSearch. Let’s start a new chat. I’m sending the same electric vehicle prompt I used for ChatGPT. It’s starting to think. Sometimes it takes a while. It can take 5 or 6 minutes and then error out. You just have to try again. ChatGPT-4o has hit 20 sources and is still going. Switching to Grok.

Grok should finish faster. It’s scanned 21 pages so far. Let them keep working. I’m moving on to Perplexity. Think of Perplexity as a search engine alternative to Google. Since it’s AI-driven, it doesn't just list links. It researches and summarizes sources like Grok and ChatGPT. Perplexity is basically the pioneer of AI-powered search. I’ll paste the same prompt here. Here’s the thing. I used a Turkish prompt. Perplexity is searching Turkish sources. It’s pulling from sites like Donanım Haber and Skywell. ChatGPT and Grok are sticking to international sources.

This is important to know. When I prompt in Turkish, Perplexity only checks Turkish pages. It returns a result based on that. As you can see, it gave me an answer very quickly. It pulled from sources like automotive sites and Skywell Turkey. It says 2025 EVs are trending due to tech and infrastructure gains. But some limitations still remain. It covers battery tech. It mentions 1,000 km ranges, production costs, and recycling. New models hit 500-800 km. Luxury models are breaking the 1,000 km barrier.

Charging times are fast. You get 80% charge in 15-20 minutes at fast stations. Home charging still takes 6-8 hours. I got exactly what I asked for. It supports the output with sources. You can click these links to verify the info yourself. It’s great for finding sources. I use Perplexity a lot. I highly recommend it. For deeper research, prompt in English. It will significantly boost your source variety. ChatGPT-4o is still running. Let's check Grok. Grok is finished. It provided a very detailed result. It scanned 82 web pages. Think about it. You have a specific question. You need an answer for it.

Without AI, scanning 82 pages is impossible. It would be a huge waste of time. Usually, you just check 2-3 sites. You grab whatever info you find, right or wrong. Now, it gives you all the sources. It extracts the main answer. Just like Perplexity did earlier. It shows battery life is 8-10 years. Or 100,000 to 200,000 kilometers. Costs are dropping too. Range varies between 200 and 500 kilometers. Some models hit 512 kilometers. Turkey has about 25,000 charging sockets as of March 2025. It says this might not be enough for long trips.

Maintenance costs are 50% lower than gas cars. You can see environmental impacts step by step. It provides performance data and main sources. Grok is quite good here. It is very usable. The research took 1 minute and 14 seconds. ChatGPT is still running, by the way. It checked 24 sources. It is hitting major outlets like The Guardian. Let it keep going. That was the search part. Finally, I want to get to NotebookLM. NotebookLM isn't just for students. It helps you master any topic. It analyzes sources and presents them in different ways. It’s a great tool for deep learning.

Upload a YouTube video or text. It generates summaries, Q&As, and timelines. These outputs help you grasp the material much faster. Take a long Wikipedia article on the Industrial Revolution. Instead of reading it, just drop the link into NotebookLM. Hit insert. It grabs the text and summarizes it. You can generate FAQs from the text. You can also create briefing docs. The podcast feature is the most impressive part. It only works in English for now. I got English results. But it should output in Turkish if your Google account is set to Turkish. There’s no real language barrier. Only the podcast is currently English-only. That will likely change soon.

I expect full Turkish support soon. The podcast feature is incredibly detailed. Two AI voices discuss the topic. It simulates a conversation to help you understand the subject better. This is another great feature. You can even analyze a YouTube video here. For example, I just paste the YouTube link. It grabs the transcript and summarizes it for you instantly. Just like with Wikipedia, you can generate FAQs, briefs, or study guides. You can also ask it specific questions about the content. This is a super powerful tool.

Let’s check if ChatGPT is done. Not yet. ChatGPT is slow. We get it. But I know the output will be high quality. That’s how I handle my daily workflow. I use multiple models like this. It really boosts my productivity. It speeds everything up. My research time is way down. Debugging code and fixing errors is so much faster with AI now. Let’s wrap up the video here. Thanks for watching. Subscribe to the SEO channel for more SEO Talk videos. Follow us for content on AI, SEO, analytics, and PPC.

Take care.

This transcript was generated using AI speech recognition and verified against subtitle style rules.

İren Saltalı

Content author

İren Saltalı

author

Contact us

How we handle your data

Open as a full page