GenAI for Business
Home › Book › Large language models and tools
Chapter 4

Large language models and tools

From GenAI for Business (2026 Third Edition) by Shubin Yu · Open in the interactive reader · Download the full PDF

Learning Objectives

After this chapter, you should be able to:

Large Language Models (LLMs) have changed how we work with natural language. Trained on enormous datasets, they can draft a memo, debug a spreadsheet formula, or work through a problem in a specialized field, all from a plain English request. This chapter compares some of the most notable LLMs currently available, including OpenAI's GPT series, Anthropic's Claude, Google's Gemini series, and challengers such as DeepSeek, Moonshot AI's Kimi, and Grok. A note before we begin: model names and version numbers churn quarterly, so treat the specific models named here as a snapshot (mid-2026) and the selection criteria as the durable lesson. We will also look at integrated platforms, playgrounds for experimentation, options for running models locally with LM Studio and Ollama, and, in Section 4.7, a method for choosing between open and closed models that outlasts any snapshot.

4.1 Leading Commercial Models: OpenAI, Anthropic, and Google

4.1.1 OpenAI's GPT Series: Pushing the Boundaries

OpenAI's GPT series sits at the front of the commercial pack. Each generation has brought improvements in architecture and capability, and the current family handles contextual understanding, multimodal processing (text, image, audio), and nuanced language generation better than anything the company shipped before. The flagship generation at press time (mid-2026) is GPT-5.6, released in three sizes named for celestial bodies: Sol, the frontier flagship and OpenAI's strongest coding and science model to date; Terra, the mid-tier workhorse; and Luna, the small, fast, inexpensive option. The generation's headline improvement is token efficiency: OpenAI reports Sol completing coding tasks with roughly half the tokens of its predecessors, which translates directly into lower cost and latency, and it ships with OpenAI's most advanced safety stack.

Around the models, OpenAI has built a serious agent platform: the Responses API (which replaced the older Chat Completions interface), the Agents SDK and AgentKit for building multi-agent systems with sandboxed execution and tracing, and Codex, its agentic coding product, which passed five million weekly active users in 2026. For business buyers, this platform layer increasingly matters as much as the model itself.

Choosing within the GPT series usually comes down to weighing raw capability (Sol) against efficiency and cost (Terra and Luna): route the hard, high-stakes work to the flagship and the high-volume routine work to the smaller tiers.

4.1.2 Anthropic's Claude Series: Focus on Enterprise and Safety

Anthropic's Claude series puts its emphasis on safe, coherent, and reliable output, and in June 2026 it gained a new top tier. Claude Fable 5, the first model in the Claude 5 family, belongs to a new "Mythos-class" capability tier that sits above the familiar Opus class. Its sibling, Claude Mythos 5, shares the same underlying model but is available only to approved organizations (such as vetted cyberdefense and biomedical research groups); Fable 5 is the generally available version, with additional safeguards around dual-use capabilities in areas like offensive cybersecurity and biology. The two-model release is itself a lesson in AI governance: as frontier capability grows, labs are beginning to segment access by risk rather than releasing one model to everyone.

Beneath the flagship sit the familiar tiers: Claude Opus, the everyday frontier workhorse; Claude Sonnet, balancing near-Opus capability with markedly lower cost; and Claude Haiku, covering fast, cost-sensitive, high-volume use cases such as sub-agent fleets. The series excels at agentic coding, long-horizon autonomous work, and writing; Anthropic trains its models with Constitutional AI to reduce harmful responses and stay aligned with user intent. Claude offers a very large context window (up to 1M tokens), native tool use, and computer use, and its Claude Code product has become one of the reference harnesses for agentic software work. Capability at the top of the family is striking: early Fable 5 deployments include a payment company's one-day migration of a 50-million-line codebase, work previously estimated in team-months. That combination suits enterprise applications, professional services, and customer-facing settings where dependability matters more than novelty, and by mid-2026 Anthropic held the largest share of enterprise LLM API spending.

4.1.3 Google's Gemini Series: Multimodality and Research Integration

Google's Gemini series was built multimodal from the start. The current line spans a Pro tier and an efficiency-focused Flash tier, with a "Deep Think" mode delivering Olympiad-level science and mathematics reasoning for premium subscribers, and the models reason across text, images, video, and audio inputs with context windows of up to one million tokens, which allows deep analysis of long documents, codebases, or multimedia content. Flash targets efficiency, offering high performance at lower cost, while the Pro tier handles the most demanding reasoning and multimodal workloads. One cautionary note for buyers: Google's next flagship slipped repeatedly through 2026 amid reliability rework, a useful reminder that roadmap promises are not products.

Google also puts its models to work in specialized research tools. NotebookLM acts as a personalized research assistant, grounding its responses in source materials you provide, which makes its answers easier to verify. Google is also exploring concepts like the "AI Co-scientist," which puts AI to work directly in the research process, from hypothesis generation to data analysis (https://research.google/ blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/).

4.2 Other Notable Models and Platforms

The "big three" do not have the field to themselves, and by mid-2026 the most important challengers are open-weight models, several of them from Chinese labs, that trail the closed frontier by months rather than years.

4.2.1 DeepSeek: Open-Weight Powerhouse

DeepSeek made its name in early 2025 when its R1 reasoning model matched frontier-level performance at a fraction of the training cost, briefly shaking global technology markets. Its current generation, MIT-licensed, with a million-token context and near-frontier coding scores, plus an ultra-cheap Flash variant, continues the pattern: open weights, efficient Mixture-of-Experts architectures, and prices an order of magnitude below the closed labs. That openness is the draw: researchers and enterprises who want customization, transparency, on-premises deployment, or lower operating expenses can simply take the weights and run.

4.2.2 Grok: Real-time Information and Conversational Style

Grok, developed by xAI (folded into the SpaceX orbit in 2026), sells itself on a conversational style often described as witty or rebellious, and on access to real-time information through its integration with the X platform. Recent releases have repositioned the line as a serious coding and agents model rather than a personality act. If your work involves tracking current events, trends, or public discourse, the real-time access is worth something.

4.2.3 Kimi and the Open-Weight Wave

The clearest sign that open-weight models have reached the frontier came in July 2026, when Moonshot AI released Kimi K3: a 2.8-trillion-parameter Mixture-of-Experts model, the largest ever published with open weights, with a 1M-token context window, native vision, and an always-on reasoning mode. On independent leaderboards it ranks just behind the top closed flagships, meaning any company can now download, inspect, fine-tune, and self-host a model within striking distance of the best closed systems, at API prices well below them.

K3 is the crest of a broader wave. Zhipu's GLM line (MIT-licensed) and MiniMax's models (the first to combine open weights, frontier coding, and native multimodality) push the same frontier, while Alibaba's Qwen line dominates the small-model tier that runs on ordinary hardware. For strategy, the implication is twofold. First, the floor price of intelligence keeps collapsing, which changes build-versus-buy math for every AI product. Second, data sovereignty is now compatible with capability: a bank or hospital that cannot send data to a US cloud can self-host a near-frontier model. The trade-offs are governance and support: open weights mean you own the deployment, the security patching, and the compliance story yourself.

4.3 Integrated platforms

Some platforms bundle several LLMs behind a single interface, so you can switch models without switching accounts. Why would you want that? Because no single model is best at everything, and comparing two answers to the same prompt teaches you more about a model's blind spots than any benchmark table.

4.3.1 Poe

Poe, developed by Quora, gives you one place to reach multiple AI chatbots, including the latest GPT, Claude, Gemini, and open-weight models. You ask a question, get an answer, and carry on the conversation, switching models whenever one disappoints you.

4.3.2 Coze

Coze is a chatbot development platform for people who may never have written a line of code. Users build customized chatbots and deploy them across social platforms and messaging apps. Its focus is personalization: the bots adapt to user-specific preferences and learn from repeated interactions.

4.3.3 Perplexity AI

Perplexity AI is a conversational search engine. It pairs large language models with web search and returns concise answers backed by cited sources, so you get a direct response to your question rather than a page of links.

4.3.4 OpenRouter

OpenRouter is the developer-side version of the same idea: a single API gateway that routes requests to hundreds of models across every major provider, closed and open-weight alike, with unified billing and automatic fallbacks. For teams, it is the fastest way to A/B test models in a real application, and its public usage rankings double as a live census of which models developers actually choose when they are paying.

4.3.5 You.com

You.com is a search engine built around a chat-first AI assistant, with conventional web search still available underneath. Its AI Modes let users switch between leading models from OpenAI, Anthropic, Google, and the open-weight ecosystem, and the company has increasingly pivoted toward enterprise research-agent products.

The pattern across all five is the same: one interface, many models, and the freedom to pick whichever fits the task in front of you.

4.4 Playgrounds

Before you commit a budget to any model, test it. Playgrounds exist for exactly this. Platforms such as OpenAI Playground, Google AI Studio, NVIDIA AI Playground, IBM watsonx.ai, and Hugging Face give developers, researchers, and businesses a low-stakes place to try the leading open-weight models (Llama, Kimi, DeepSeek, Qwen) and others: experiment with prompts, explore fine-tuning, check integration options. Most offer free tiers or trials, so a half-day of testing costs you nothing but the half day. In my experience that half day saves weeks of regret later.

4.4.1 Arena (formerly LMArena)

Arena (https://arena.ai/), the platform long known as Chatbot Arena and then LMArena, is the largest blind human-preference evaluation in the field: you type a prompt, two anonymous models answer, you vote, and millions of such votes (more than seven million across several hundred models by 2026) feed public leaderboards. It is no longer a single ranking. Separate leaderboards now cover text, vision, coding, web development, search, image, and video generation, because the best model depends entirely on the task, and you should look at the category that matches yours. It is open access, and it gives you a transparent way to work out which model suits your needs, whether that is conversational fluency, response accuracy, or domain-specific expertise.

4.4.2 Nvidia build platform

Nvidia's LLM comparison platform (https://build.nvidia.com/) is another useful benchmarking stop. Built on Nvidia's hardware acceleration expertise, it lets you assess different LLMs for speed and accuracy across various tasks.

4.4.3 Google AI Studio

Google AI Studio (https://aistudio.google.com/) is the venue for comparing Google's own LLMs, such as Gemini, against other available models. Beyond benchmarks, it shows how a model behaves under awkward real-world conditions: limited compute, or a genuinely complicated user query.

4.4.4 Hugging Face

Hugging Face hosts a wide array of AI models, including the Llama, DeepSeek, Kimi, and Qwen families, many testable through the platform's interface. Many models are free to try; subscription plans add features on top.

Each of these platforms has its own strengths and access levels, but any of them will let you kick the tires before you buy.

4.4.5 OpenAI Playground

The OpenAI Playground is OpenAI's own web interface for experimenting with its language models, including the current GPT family in all its sizes and earlier models. Type a prompt, watch the response come back in real time. Free access comes with usage limitations; extended access may require a subscription.

4.4.6 IBM watsonx.ai

IBM's watsonx.ai is an enterprise-ready AI studio offering a selection of foundation models, including IBM's own Granite series alongside other open-source models. You can build, train, and deploy models on the platform, and a free trial lets you explore before committing.

4.5 Local Deployment of Language Models

Not everyone can, or should, pay for a frontier model over an API. Open-weight models let you work with capable LLMs at no per-token cost, on your own hardware, with trade-offs in sophistication, convenience, and the operational burden you take on.

For individuals or institutions that put data security and customization first, running language models locally is an attractive option. Two tools dominate: LM Studio, a desktop application with a graphical interface, and Ollama, a lightweight command-line runtime that has become the standard way to serve local models to other applications (MimiWork, Chapter 16, and most agentic tools can point at it directly). With either, nothing you type ever leaves the building. For healthcare, finance, legal work, or any team handling personal data, that alone can settle the question. Be realistic about hardware, though: the frontier open models (the largest Kimi, DeepSeek, GLM, and Llama variants) need server-class GPUs, while the small-model tier runs comfortably on a well-equipped laptop with 16 to 32 GB of memory.

Which model to download. For most readers in 2026 the answer is a member of Alibaba's Qwen family, which dominates the small-model tier: pick the largest variant that fits your memory, typically in the 4B to 30B parameter range, and choose an instruction-tuned version. DeepSeek's distilled and smaller variants are the strongest choice when reasoning and coding matter more than general chat. Google's Gemma, Mistral Small, and OpenAI's gpt-oss releases are solid alternatives with permissive licenses, and Mistral's larger Apache-licensed models remain the European option for organizations with EU sovereignty requirements. None of these rival the closed flagships in depth, but for summarization, drafting, extraction, classification, and private prototyping, the gap barely shows, and they cost nothing per token, which makes them especially useful for classrooms, non-profits, and privacy-critical pilots.

4.5.1 Open Weights: The Case For and Against, and Why It Matters

Open-weight models deserve a balanced verdict, because the decision to rely on them is a strategic one. The case for is strong. They cost nothing to license and little to run at the small end. They can be inspected, fine-tuned, and modified, which closed models cannot. They run wherever you put them, including air-gapped environments, sovereign clouds, and laptops, so data never crosses a border or a vendor boundary you did not choose. They cannot be withdrawn, repriced, or silently changed by a vendor, which means an eval suite that passes today passes next year. And they are now close enough to the frontier, as Section 4.2.3 showed, that for most business workloads the difference is not decisive. The case against is equally real. Open weights transfer the entire operational burden to you: hosting, scaling, security patching, monitoring, and the compliance story are yours, and a small team can find that burden larger than the API bill it replaced. The most capable open models are large and expensive to serve well. Licenses vary and some restrict commercial use. Support is a community, not a contract. And "open weights" is not "open source" in the traditional sense: the training data and process usually remain private, so transparency has limits.

Why this matters more every year is the argument of Section 13.6. AI models are becoming infrastructure, as fundamental to how organizations operate as databases, networks, and electricity. Infrastructure that a single foreign vendor can reprice, restrict, or switch off is a strategic vulnerability, and every government and large enterprise has now noticed. Open-weight models are the mechanism by which capability becomes a commodity that anyone can hold rather than a service that a few companies rent out, and their existence disciplines the pricing and behavior of the closed labs even for customers who never run one. A sensible organization in 2026 does not choose between open and closed. It uses the closed frontier where capability is decisive, keeps an open-weight fallback that its eval suite has already qualified, and builds its scaffolding so that switching is a configuration change rather than a rewrite. Section 4.7 turns that stance into a selection method.

4.5.2 Tutorial: LM Studio

To install and run a local model in LM Studio, follow these steps. We use a small open-weight model that genuinely runs on a well-equipped laptop:

  1. Download LM Studio: visit lmstudio.ai and download the installer for your operating system (macOS, Windows, or Linux).
  2. Install and launch it.
  3. Download a model:
    • Open the 'Discover' (search) tab.
    • Search for "Qwen", "DeepSeek", "Gemma", or "gpt-oss".
    • Select a variant sized for your machine (roughly 4B to 20B parameters for most laptops; LM Studio flags models that will not fit).
    • Click 'Download'.
  4. Load the model:
    • Go to the 'Chat' section and choose the downloaded model from the loader at the top.
    • Start chatting. Nothing leaves your machine.
  5. Optional: open the 'Developer' tab and start the local server. LM Studio then exposes an OpenAI-compatible API on your machine, so any application that speaks the OpenAI API, including the agentic tools in Chapter 16, can use your local model.

4.5.3 Tutorial: Ollama

Ollama trades the graphical interface for simplicity and automation, and it is the runtime most agentic tools expect when you say "use a local model":

  1. Download Ollama from ollama.com and run the installer (macOS, Windows, or Linux). It installs a background service and a command-line tool.
  2. Open a terminal and pull a model, for example ollama pull qwen3 or ollama pull deepseek-r1. The library at ollama.com lists sizes; the default tag is usually a laptop-friendly variant, and you can request a specific size such as qwen3:8b.
  3. Run it: ollama run qwen3 opens an interactive chat in the terminal. Type your prompt; type /bye to exit.
  4. Use it from other applications. Ollama serves an OpenAI-compatible API on your machine at port 11434 by default. In MimiWork, choose Ollama under Settings, Models; in OpenCode, Kimi Code, or Claude Code, point the model provider at the same local address. Your agent now runs entirely on your hardware.
  5. List and remove models with ollama list and ollama rm; models are large, and a laptop fills up quickly.

Check your hardware first with either tool. Large models are memory-hungry, and an underpowered machine will make the experience miserable. Start small, confirm the workflow, then scale the model up until quality is good enough for the task and no further.

4.6 A Habit, Not a Snapshot: How to Check What Is Current

Every model name in this chapter is a photograph of a moving object, taken in mid-2026. The durable skill is not memorizing the lineup; it is the fifteen-minute quarterly check that keeps your mental map current. Three stops suffice. Arena (arena.ai, formerly LMArena) shows how models rank on blind human preference, by task category. Artificial Analysis (artificialanalysis.ai) tracks independent benchmarks, prices, and speed across every provider on one page. And each vendor's own pricing page is the only authoritative source for what you will actually pay. When a colleague or vendor names a model, the reflex to build is simple: check when it shipped, what it costs per million tokens, and how it ranks against the incumbent, before anyone builds anything on it. Fifteen minutes, four times a year, keeps you permanently ahead of most of the room.

4.7 How to Choose a Model: Open or Closed, and Which One

Everything in this chapter converges on one decision that every team faces repeatedly: which model, for this workload, this quarter. The wrong way to decide is by brand familiarity or by whichever leaderboard a vendor quoted last. The right way, and the method that the companion pages at gaiforbusiness.com apply to their quarterly landscape update, weighs three things: capability on your task, cost at your volume, and sovereignty over your data. Every other consideration folds into one of those three.

Capability means performance on your work, not on a benchmark. Build the ten-to-fifty-case eval set that Chapter 5 describes before you compare anything, then run the candidates through it. Public leaderboards (LMArena, now simply Arena, for blind human preference; Artificial Analysis for independent benchmarks, speed, and price) are for shortlisting, and their categories matter: a model that leads in coding may trail in long-document analysis or in your language. Check the context window your documents actually need, whether the model handles the modalities you use, and, for agentic work, how well it follows tools and instructions over long tasks, which is where the frontier models still separate from the pack.

Cost is price per million tokens multiplied by your real volume, adjusted by the three levers Chapter 5 explains: prompt caching, batching, and routing. Routing is the big one. Almost no workload needs the flagship for every call; the standard pattern is a small, cheap model for the high-volume routine steps and a frontier model for the few steps where judgment matters, with your eval suite deciding which is which. Price the total system, including the engineering and eval work, not the token rate on a pricing page. And remember that open-weight models trade token cost for hosting cost: free per token is not free per month once you count GPUs, staff, and on-call.

Sovereignty asks who can see your data, who can change or withdraw the model, and under which jurisdiction both sit. Consumer chat tiers are for public information only. API and enterprise agreements with training exclusion and regional hosting cover most business data. Regulated or genuinely sensitive data, and any workload where continuity is a strategic requirement, calls for either a cloud-marketplace deployment in your own region (Bedrock, Azure, Vertex) or an open-weight model you host, and Section 13.6 explains why this consideration is growing in weight every year.

Put together, the decision is usually not a single winner but a portfolio. Use a closed frontier model where capability is decisive and the data permits it. Use an open-weight model where data cannot leave your control, where cost at scale dominates, or where you need a fallback that no vendor can take away. Route between them with your own evals as the referee. Then run the fifteen-minute quarterly check from Section 4.6, re-run the eval suite whenever a serious new model ships, and switch when the numbers, not the headlines, say so. A team that works this way is never captured by a vendor and never surprised by the landscape, which is the most that any snapshot of the model market can promise.

Discussion Questions

  1. For your most sensitive workload, argue both sides: closed frontier API versus self-hosted open-weight model. Which wins, and on what grounds?
  2. What concrete signals would make you switch primary model vendors, and how would your eval suite tell you it is time?
  3. Run the fifteen-minute currency check now. What has changed since this book's mid-2026 snapshot, and does any of it change a decision?
This chapter is part of GenAI for Business, free to read in full. Continue with the next chapter, browse the glossary, or use the free templates it references.
← Communicating with GenAIAPI integration →