AI Services and Tools (1): Chat Services (ChatGPT / Gemini / Claude / Perplexity / Open-Weight Models)#

There are organizations that build AI models. They publish their models, run their own services on top of them, and offer APIs so developers can use the models. Others take those published models and build “derivative models” optimized for their own purposes, then publish them again or offer them as services or APIs.

This chapter covers the chat services that the largest number of people use every day. The list of “what you can do” with each service is handled all at once in Part 4, “Boosting Work Productivity,” so here I focus on which organization builds each product and in what direction, and on how the services differ from one another.

First, Know This: Features Are Converging#

A few years ago there were many things that “ChatGPT can do but the others can’t.” Today the major services offer almost the same bundle of features. Web search, file upload and analysis, image generation and editing, voice conversation, long-document handling, building your own chatbot (under names like GPTs, Gem, and Project), memory that carries across conversations, and even an agent mode that carries out multi-step tasks on its own are now mostly common to all of them.

So the question “which service is absolutely better” is losing its meaning, and the deciding factors have become relative strengths and fit with the ecosystem you already use. The conventional wisdom at the time of writing is summarized below. Rankings flip every time a model generation changes, so read this as a starting point rather than a conclusion.

ChatGPT (OpenAI)#

ChatGPT is a chat-based AI service that OpenAI launched on November 30, 2022. It reached one million sign-ups within five days of launch, and some analyses put its monthly user count at around one hundred million within two months. This showed that, compared with earlier AI services, ChatGPT had reached “a level of polish that ordinary users could use right away,” and global interest in AI services rose sharply afterward.

The GPT models behind ChatGPT have advanced in three directions across generations. The first is multimodality. As the models came to handle not just text but voice and images naturally as input and output, experiences like low-latency voice conversation became possible. The second is reasoning. A mode was introduced in which the model goes through an internal thinking step before answering, raising accuracy on complex problems. The third is agentic work. This is the push to raise “the quality you get when you hand over a complex job from start to finish,” across professional tasks like building spreadsheets or presentations, coding, understanding long documents, and tool calling.

My own habits have changed too. Where Google used to be my first option for search, I am now closer to “get an overview from ChatGPT first, then check the original sources when needed.” That said, AI models still hallucinate, so one thing has not changed: information that feeds important decisions, or legal or financial judgments, needs source checking and cross-verification.

GPTs: Building Your Own ChatGPT#

GPTs are “customized versions” of ChatGPT. When you build your own GPT, you typically combine three elements.

• Instructions: rules such as tone, role, prohibitions, and output format
• Knowledge: having it answer with reference to documents you upload (a lightweight version of the RAG we saw in Chapter 2)
• Tools (Actions): calling external APIs when needed and folding the results into the answer (a form of Function Calling)

Calling external APIs requires authentication (security) settings and a request / response interface definition (an OpenAPI schema). In other words, you have to spell out “which API I provide and in what format” before a GPT can use that tool safely.

You can find a GPT that fits your needs in Explore GPTs, but it is hard to say that GPTs have been “explosively” adopted by the general public. Purpose-built AI services and tools appeared quickly, and once MCP became the standard, the path for “bringing my service inside an AI” widened from GPTs alone to several routes. Still, it remains one of the lightest ways to put your business in front of ChatGPT users.

Beyond Chat#

OpenAI has also expanded its products beyond the chat service. Products like the video generation service Sora, the coding agent Codex, and the AI browser Atlas share the same account and subscription as ChatGPT while shifting the weight toward “AI doing the work on your behalf.” The remaining chapters of Part 3 and the implementation chapters cover this trend.

Gemini (Google)#

The story that Google declared a “code red” and convened emergency meetings when ChatGPT launched in November 2022 is fairly well known. Google then released Bard, iterated on it quickly, and rebranded it as Gemini in February 2024. Gemini has since caught up considerably in quality and features, but more people still think of ChatGPT first when they hear “AI chatbot.”

Gemini resembles ChatGPT in many ways. People have already grown used to “using AI through chat,” so making the UI / UX or core features too different would only raise the barrier to entry. Gemini’s distinctive strengths are two. One is that it was ahead early on in handling very long context, and the other is its integration with Google Workspace (Gmail, Docs, Drive, Calendar). For an organization that already works inside the Google ecosystem, it is the lowest-friction choice.

Gem is a feature very similar to ChatGPT’s GPTs: you set instructions in advance or supply knowledge for a specific purpose and get “your own chatbot.” At the time of writing, however, Gem did not offer the ability to connect external APIs and call tools the way GPTs Actions does. Feature details like this change every few months, so check again when you use it.

From a developer’s perspective, rather than extending the Gemini app itself, the more common approach is to attach the model to your service through the Gemini API (and Vertex AI) and extend product features through tool integration.

Claude (Anthropic)#

Claude is a generative AI model and service developed by Anthropic. Anthropic is known as a company co-founded by people including former OpenAI researchers, and Claude is often introduced through its emphasis on “being helpful safely” (for example, Constitutional AI).

Anthropic refreshes Claude generation by generation, and within each generation it offers a top-tier model (the Opus line), a balanced model (the Sonnet line), and a lightweight model (the Haiku line). Officially, the models are positioned as strong at coding, agents, and computer use (official announcements). Among developers, Claude is frequently cited as especially strong at code generation, refactoring, debugging, and code review, and many users also prefer it for long-form writing and careful reasoning.

Within the Claude lineup, what a solo startup should pay the most attention to is the agent products: Claude Code, which takes on codebase-level work from the terminal; Cowork, which handles working folders and the browser for non-development tasks; and the MCP and skills ecosystem they share. I run this blog’s writing, translation, and news-brief production as procedures (skills) on top of Claude Code, and the implementation chapters cover that approach in detail.

Perplexity#

Perplexity is a search / research service in the form of “AI does the searching for you and shows its sources alongside the answer” (perplexity.ai). Rather than handing you a list of links like ordinary search, it works out the intent of your question and presents a summarized answer first (narrowing the scope with follow-up questions when needed), together with links to the web documents it drew on. That makes it especially useful when you need to “get a quick overview and follow the sources to verify,” such as market / competitor research, checking policies or terminology, and catching up on current issues. Sources attached to an answer do not guarantee that the answer is correct, though, so for anything important, make a habit of checking the originals yourself.

Now that ChatGPT, Gemini, and Claude all have web search, Perplexity’s differentiation has narrowed, but as a product whose “core business is search,” it is still widely regarded as ahead in the polish of its source attribution and research flow.

Open-Weight Models (Llama, DeepSeek, Qwen, and Others)#

All of the services above run the model on the company’s servers, and users receive only the results. There are also many models that go the other way: models whose weights are published so anyone can download and run them. Meta’s Llama opened this trend, and DeepSeek, Alibaba’s Qwen, Mistral, and other organizations have joined in. They are commonly called “open source,” but cases where the training data and code are fully published are rare, so open-weight models is the more accurate term. License terms (whether commercial use is allowed, user-count limits, and so on) also vary by model, so check before using one in a business.

The strength of open-weight models is not that “they are free to use” but that they give you the option of running AI in your own environment without sending data outside. You can download a model from Hugging Face and run it on a local PC with a tool like Ollama, or call it on a managed cloud platform (AWS Bedrock, Google Vertex AI, and others). An AI model here is not a single executable but closer to a huge function that computes “input → output,” so you also need a runtime that loads the model, processes tokens, and wires up input and output.

Running a model yourself, however, brings hardware and operational burdens, and depending on the task there is still a quality gap compared with top-tier commercial models. For a solo startup, the realistic sequence is usually to start with a commercial API and look at open-weight models when security requirements or cost structure make it necessary. That decision continues in the next chapter, “APIs and Model Selection.”

Summary#

• The features of the major chat services have converged, and the deciding factors are relative strengths and ecosystem fit.
• By conventional wisdom, ChatGPT has the broadest general-purpose coverage, Gemini has long context and the Google ecosystem, Claude has coding, writing, and agents, and Perplexity has source-backed research.
• Open-weight models give you options for data sovereignty and cost, but they come with operational burden.
• What matters most to a solo founder is not the features inside the chat window but the agent products and APIs each service offers. That story begins in the next chapter.