AI Services and Tools (1)#
There are organizations that build AI models. They may publish their models, offer their own end-user services powered by those models, or provide APIs so developers can use the models. Based on public models, others may also create and release “derivative models” optimized for their own purposes, or ship them as services/APIs.
Meanwhile, many teams don’t train models from scratch. Instead, they integrate AI features into products quickly by using APIs offered by others. In this ecosystem, new services and tools keep appearing, helping organizations of all sizes—from startups to large enterprises—apply AI to business. In this chapter, we’ll focus on well-known AI services and tools and introduce what each offers and how they’re commonly used.
ChatGPT (made by OpenAI)#
ChatGPT is a chat-based AI service released by OpenAI on November 30, 2022. It reportedly reached one million users within five days, and some analyses claimed it hit the scale of 100 million monthly active users within two months. This suggests ChatGPT delivered a level of “polish that regular users could actually use right away,” and after its launch, global interest in AI services surged.
ChatGPT initially used GPT-3.5–class models, and later GPT-4 became available primarily through paid plans. In 2024, OpenAI announced GPT-4o, where “o” stands for “omni,” emphasizing more natural handling of multimodal inputs/outputs like text, voice, and images.
The key point is that it’s not only “the model got smarter,” but that the conversation experience became much smoother. For example, if voice chat responses lag by seconds, the conversation feels awkward. As processing paths improved (or multi-step pipelines became more integrated), perceived latency dropped, making real-time conversational use much easier. If you watch OpenAI’s GPT-4o launch stream, they keep the entire demo as a spoken conversation—something that would have been hard to showcase when responses routinely took several seconds.
OpenAI also announced GPT‑5.2 on December 11, 2025 (official post). OpenAI describes GPT‑5.2 as improving performance for professional knowledge work (e.g., spreadsheets and presentations), coding, understanding long documents (long context), and tool use (agent-like tasks), and says that hallucination errors decreased in real-world usage. In other words, beyond “more natural conversation,” the direction is better end-to-end quality and reliability when you delegate complex tasks.
Now let’s look at common ways to use ChatGPT.
Information search#
Even when you can’t think of good search keywords, ChatGPT can help if you describe what you’re looking for in everyday language, as if you were asking a friend. Early on, it struggled with events after its training cutoff (latest news, new product info, etc.), but over time it added features like web browsing/search, moving toward answers that can reference more up-to-date information.
After ChatGPT launched, stories about “search being threatened” (e.g., Google’s “code red” anecdotes) became widely known. Personally, I used to default to Google search first; now I’m closer to “summarize with ChatGPT first, then check primary sources as needed.”
However, because AI models can hallucinate, for information used in important decisions—business, legal, or financial—you should verify sources and cross-check.
Summarization, translation, writing, and editing#
ChatGPT can summarize long documents or articles to extract key points, translate across languages while preserving context, draft writing on a requested topic, and help review and edit text you’ve written.
Writing and reviewing code#
Developers use ChatGPT to work more efficiently on writing and reviewing code. If you describe requirements for a class or function, you can get code plus explanations. If you paste code and ask for review, it can suggest fixes and improvements.
Even if you’re not a developer, you can still use ChatGPT to apply simple automation at work. For example: “Write Python code that visits https://some.ecommerce.site/some/product/id and extracts the product name and price.” Then: “I want to run this Python script daily and save the price to an Excel file by date. I’m not a developer and I’ve never run Python before. Explain step-by-step what I should do.” You can get a guided setup plan that walks you through the whole process.
Data analysis and visualization#
ChatGPT can take data in various formats, analyze it, and produce different kinds of outputs. You can input data as plain text, upload files like spreadsheets, or select Google Workspace files such as Google Sheets. Outputs can be written analysis or charts. For example, you can attach a Google Sheet that tracks ongoing and planned tasks and ask: “For both ‘ongoing’ and ‘planned’ tasks, aggregate ‘development duration’ by ‘category’ and draw a bar chart of total development duration by category.”
Image generation#
You can generate images with ChatGPT and also select and edit parts of a generated image. You can create artwork, app icons, game assets, and more. One limitation is that it can still be hard to control fine-grained details exactly as intended. It can feel like working with a brilliant but unpredictable painter. It tends to fit situations where you set a direction and iterate based on what comes out, rather than cases where every detail must match a fixed spec. Another good approach is using generated results as inspiration, then refining or recreating them separately.
Video generation#
It’s not quite accurate to say ChatGPT itself has video generation “fully integrated,” but you can create videos through OpenAI’s video generation service Sora. Sora can generate short videos from text prompts, and also supports workflows like changing style or modifying parts of a scene based on images or existing video inputs (remix/editing). It’s well suited to iterative generation: try, review, and refine until you get the direction you want.
Ideation#
Creativity used to be framed as uniquely human. But generative AI models, which learn patterns in a way loosely analogous to how humans learn and can “generate” outputs, can now help people produce creative results. With ChatGPT, you can brainstorm ideas for new projects, and it’s also very helpful for naming and copywriting.
Problem solving#
All of the use cases above ultimately fall under “problem solving.” Beyond that, you can use ChatGPT to tackle many kinds of problems by describing the situation in detail and asking it to propose solutions.
GPTs#
GPTs are essentially “customized versions” of ChatGPT. When you create your own GPT, you typically combine:
- Instructions: rules like tone, role, constraints, and output format
- Knowledge: documents/materials you upload so it answers based on them (a form of “document-grounded answering”)
- Tools (Actions): calling external APIs when needed and incorporating the results in the response (= a form of function/tool calling)
As discussed in the RAG section, an LLM answers based on what it already learned during training, so for internal documents or the latest policies—information the model doesn’t inherently know—it’s often better to attach documents and ground the answer. GPTs make it relatively easy to set up both “document-grounded answering” and “tool calling,” which greatly expands what you can do with ChatGPT.
When building your GPT, you can use the “Create” tab to build it by chatting with the GPT Builder, or you can configure it directly in the “Configure” tab. Builder is fine for simple GPTs, but if you need to call external APIs, you usually add Actions in the Configure tab.
Because calling external APIs means integrating with other services, you need authentication/security settings and a clear request/response interface definition (e.g., an OpenAPI schema). In other words, you must specify “what API I provide and in what format” so the GPT can use the tool safely.
You can browse GPTs from “Explore GPTs,” but it’s still not easy to find many GPTs that stand out as “high quality.” Many people expected it would naturally improve over time, but in practice the market has also diversified into many purpose-built AI services and tools, so it’s hard to say GPTs have “exploded” in mainstream usage.
Perhaps the most interesting direction for many readers is using GPTs to extend your business through ChatGPT. For example, your GPT could post content directly to a blog or social network, book a hotel or flight, or integrate another independent AI service as part of your workflow.
Gemini (made by Google)#
After ChatGPT launched in November 2022, stories spread that Google declared “code red” and held emergency meetings. Google released Bard and iterated quickly, then rebranded it as Gemini in February 2024. As of January 2026, Gemini has caught up significantly in quality and features, but many people still think of ChatGPT first when they hear “AI chatbot.”
In many ways, Gemini is similar to ChatGPT. People have become used to “using AI through chat,” and if a product’s UI/UX and core flows are too different, it can raise the barrier to entry. So Gemini’s use cases largely overlap with ChatGPT’s: information search, summarization, translation, writing, editing, code writing/review, data analysis, ideation, and problem solving. It also supports multimodal features like file upload/analysis, image generation/editing, and video generation. (Availability can vary by country, account, and plan.)
Gems (custom) are very similar to ChatGPT’s GPTs. You can create “your own chatbot” for a specific purpose by predefining instructions and providing knowledge. Like GPTs, this reduces repeated prompting by letting you reuse a consistent set of requirements, tone, and reference materials. However, as of January 2026, Gems do not provide a feature like GPTs Actions where you can connect external APIs to run tasks (task execution) or call tools.
From a developer’s perspective, rather than extending the Gemini consumer app itself, it’s more common to integrate Gemini via the Gemini API (and platforms like Vertex AI) and expand product capabilities through function/tool calling integrations. It also aligns with Google’s direction that it ran the Gemini API Developer Competition in 2024 with substantial prizes.
Claude (made by Anthropic)#
Claude is a generative AI model/service developed by Anthropic. Anthropic is known as a company co-founded by people including former OpenAI researchers, and Claude is often described in the direction of producing “helpful answers safely” (e.g., Constitutional AI).
As of January 2026, Anthropic released the Claude 4.5 series, and Claude Opus 4.5 is positioned as the top-tier model that Anthropic officially describes as “best for coding, agents, and computer use” (official announcement). Like ChatGPT and Gemini, Claude offers a chat experience (claude.ai), and among developers it’s frequently mentioned as especially strong for code generation, refactoring, debugging, and code review.
Perplexity#
Perplexity is a search/research service where “AI searches on your behalf and shows sources alongside the answer” (perplexity.ai). Instead of returning only a list of links like traditional search, it first presents a synthesized answer that reflects your intent (often narrowing the scope through follow-up questions), and includes links to the web pages it referenced. This is especially useful when you need to quickly build an overview of a topic and then follow sources—for example, market/competitor research, policy/terminology checks, or tracking recent issues. Still, even with sources attached, the answer isn’t guaranteed to be correct, so for important matters you should open and verify the originals.
LLaMA (made by Meta)#
LLaMA (Large Language Model Meta AI) is a family of LLMs released by Meta. In many cases, it’s more accurate to think of it not as “open source,” but as an open-weight model (i.e., the model weights are available). That means you can download and run the model (within license terms) and fine-tune it if needed, which has made it very influential in the research/developer ecosystem.
You can also access LLaMA through chat experiences like Meta AI (meta.ai), though availability can vary by region/account. The big advantage of LLaMA is that you can “bring it and run it yourself.” You can download models from Meta’s downloads page or from the Hugging Face collection, then run them locally or on your own servers using tools like Hugging Face Transformers or Ollama. Because an AI model isn’t a single executable but more like a huge function that maps inputs to outputs, you also need a runtime/library that loads the model and handles tokenization and I/O.
The value of an open-weight model isn’t just that it can be “free.” It also gives you the option to run AI in your own environment without sending data externally (or at least minimizing it). That said, running locally can add hardware and operational burden, and open models are not automatically free from issues like hallucinations, security, copyright, and licensing—so if you use them in products, you need appropriate validation and guardrails for the use case.