The Big Three

Starting Here? Read This First

If you've jumped straight to this module hoping to pick a model and get started, you're in good company—this is exactly what most people want. But foundation models are tools, and like any tool, their value depends on how skillfully you use them. Before you dive too deep here, consider at least skimming these foundational concepts from earlier modules:

Now, let's meet the models.^[1]

The Med Student Analogy

Think of each foundation model as a brilliant medical student who has read essentially everything ever published—every textbook, every journal article, every clinical guideline, every case study, and frankly, every Reddit thread and random blog post too. This student has near-perfect recall of patterns across all that material and can synthesize information across domains in ways that would take you hours or days.

But here's what's crucial: this med student has never actually seen a patient. They haven't felt the resistance of tissue, watched a parent's face crumple at difficult news, or learned from the case that didn't follow the textbook. They know what clinical reasoning looks like on paper, but they don't have clinical judgment.

This framing helps calibrate expectations:

• They're genuinely useful for the kind of work where pattern recognition and information synthesis matter—drafting notes, summarizing literature, explaining concepts, generating differential diagnoses for discussion.
• They require supervision the same way any trainee does. You wouldn't let even a brilliant third-year student sign notes unsupervised, and you shouldn't let an AI do so either.
• They need clear instructions. Just as you'd give a student specific guidance ("I need a note that addresses the parent's concern about developmental delay, focuses on what we observed today, and includes our reasoning for not ordering imaging"), you need to give the model explicit context and constraints.
• They sometimes confabulate. A student who doesn't know something might make up an answer that sounds plausible rather than saying "I don't know." These models do the same—we call it "hallucination," but it's really just pattern-matching in the absence of actual knowledge. This is why you verify before you trust.

With that framing in mind, let's look at who these three "students" are and what each brings to your practice.

Knowing Your Options

Before diving into each model, it's worth understanding the landscape:

• ChatGPT is the household name. For most people, "ChatGPT" is synonymous with AI chat. It has the largest consumer user base and the most cultural awareness. When your patients or colleagues mention "AI," they're usually thinking of ChatGPT.
• Many people don't realize they already have Gemini. If you have a Google account, you have access to Gemini. It's built into Google Search, available at gemini.google.com, and integrated throughout Google Workspace. Yet many users have never tried it.
• Claude is often discovered second. People typically find Claude when looking for alternatives, exploring options for specific use cases, or hearing recommendations from colleagues. It has a smaller but dedicated user base.

None of this tells you which is "best"—that depends entirely on your needs, your ecosystem, and your preferences. The point is simply: you have options, and it's worth knowing what they are.

ChatGPT: The First Mover

The Story

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, and others with the stated mission of developing artificial general intelligence that benefits humanity. The company initially operated as a nonprofit research lab before restructuring in 2019 to attract the investment needed for increasingly expensive AI training.

The GPT (Generative Pre-trained Transformer) architecture emerged from this research, with GPT-1 in 2018, GPT-2 in 2019 (initially withheld due to concerns about misuse), and GPT-3 in 2020. But the inflection point came on November 30, 2022, when OpenAI released ChatGPT as a free research preview. Within five days, it had a million users. Within two months, it reached 100 million—the fastest-growing consumer application in history.

That explosive growth fundamentally changed how the world understood AI. Suddenly, anyone could have a conversation with a system that felt like talking to a knowledgeable colleague. The technology wasn't new, but the accessibility was.

The Current Offering

As of early 2026, ChatGPT operates across several tiers:

What You Get

Ecosystem: The most mature AI ecosystem. The GPT Store contains customized applications for specific use cases, extensive plugin support, and integrations with tools many people already use. If you want to find a pre-built solution for a specific task, ChatGPT's ecosystem is the most likely place to find it.

Features: Advanced Voice Mode for natural conversation, image generation through DALL-E, Code Interpreter for data analysis and visualization, and web browsing for current information.

HIPAA pathway: BAAs available for API services, Enterprise/Edu plans, and the new ChatGPT for Healthcare tier deployed at major academic medical centers. ChatGPT Free, Plus, Pro, Team, and the free ChatGPT for Clinicians tier are explicitly not covered by BAAs and cannot be used with protected health information.

Claude: The Safety-First Approach

The Story

Anthropic was founded in 2021 by Dario and Daniela Amodei, along with several other former OpenAI researchers and executives. The founding team included key figures in AI safety research, and that orientation shaped the company's approach from the beginning.

The company developed "Constitutional AI," a training approach that uses AI feedback (rather than exclusively human feedback) to shape model behavior according to a set of principles. The goal was to create systems that are helpful but also harmless and honest—what Anthropic describes as the "HHH" framework.

Claude 1.0 launched in March 2023, positioning itself as a thoughtful alternative to ChatGPT. The Claude 3 family arrived in March 2024, introducing the Haiku/Sonnet/Opus tiering (small/medium/large models with different capability and cost profiles). In February 2026, Claude Opus 4.6 launched with a one-million-token context window and "agent teams" that can coordinate across shared codebases. Anthropic closed a record $30 billion funding round, valuing the company at approximately $380 billion. Later that month, Claude Sonnet 4.6 arrived—delivering near-Opus performance at one-fifth the cost, and becoming the default model for both free and Pro users. In April 2026, Claude Opus 4.7 shipped with materially stronger coding (+13% on SWE-bench), higher-resolution vision, the new /ultrareview multi-agent review workflow, and an xhigh reasoning effort level—all at the same pricing as 4.6. Claude Opus 4.8 followed on May 28, 2026—again at the same pricing—with 128K max output tokens, 88.6% on SWE-Bench Verified, a new Fast mode ($10/$50 per million tokens at ~2.5× speed), and Dynamic Workflows in Claude Code, which can orchestrate up to 1,000 subagents on a single task. On June 30, 2026, Claude Sonnet 5 arrived—a one-million-token context window, 128K max output, and the default model both in Claude Code from launch and for Free and Pro users on claude.ai. On July 24, 2026, Claude Opus 5 replaced Opus 4.8 as the current Opus, at the same $5/$25 per million tokens: a one-million-token context window, 128K max output, and a knowledge cutoff of May 2026. It is now the default model on Claude Max and the strongest model available on Pro. Opus 4.8 still works but has moved to Anthropic’s legacy list.

The Current Offering

What You Get

Ecosystem: Smaller than ChatGPT's but growing. The MCP (Model Context Protocol) allows integrations with external tools and data sources. Claude Code provides command-line AI assistance for developers. The ecosystem emphasizes depth over breadth.

Features: A one-million-token context window across the current lineup—Opus 5, Sonnet 5, and the top-end Fable 5—for working with substantial documents. Projects feature for organizing related conversations. Artifacts for generating code, documents, and other outputs. Computer use capabilities for agentic workflows. No native image generation, but can analyze images you provide.

HIPAA pathway: BAAs available through the API with Zero Data Retention (ZDR) agreement. Consumer chat interface (Claude.ai) is not covered. AWS Bedrock provides the most straightforward enterprise pathway.

Gemini: The Integration Play

The Story

Google's path to Gemini began with earlier AI efforts including LaMDA (the model behind the original Bard chatbot) and PaLM (Pathways Language Model). But Google's extensive AI research, which predates the current generation by decades, positioned the company as a natural major player once the race began.

Gemini was announced in December 2023 as Google's response to GPT-4, with the company emphasizing its native multimodal training—the model was trained from the beginning on text, images, and other modalities together, rather than having capabilities bolted on afterward.

Gemini 1.5 arrived in February 2024 with a groundbreaking 1-million-token context window (later expanded to 2 million)—dramatically larger than competitors at the time. In November 2025, Google announced Gemini 3, positioned as "our most intelligent model" with significant improvements in reasoning, multimodal understanding, and agent capabilities. In February 2026, Gemini 3.1 Pro doubled the reasoning performance of its predecessor and dominated 13 of 16 major AI benchmarks, now powering NotebookLM. At Google I/O 2026 (May 19), Google launched Gemini 3.5 Flash—rolled out to all users and roughly 4× faster than peer frontier models on agentic benchmarks—alongside Gemini Omni Flash, a new multimodal model for Plus/Pro/Ultra tiers. Gemini 3.5 Pro has still not shipped. As of late August 2026 its release has slipped more than once, and Google’s releases since—Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Flash Cyber on July 21, then Gemini 3.7 Flash on August 13—still did not include a Pro model. Every specification circulating for 3.5 Pro is unconfirmed.

The Current Offering

What You Get

Ecosystem: Deep Google Workspace integration. If your organization lives in Gmail, Docs, Sheets, and Meet, Gemini works natively within those tools—the side panel in Google Docs helps you write, Gemini in Gmail drafts responses and summarizes threads. For organizations already committed to Google, this reduces friction dramatically. If you don't use Google Workspace, much of this advantage disappears.

Features: 1-2 million token context window—the largest available—for working with enormous documents or entire codebases. Strong multimodal capabilities including video analysis. Image generation through Imagen. Gemini Live for voice interaction.

HIPAA pathway: Google Workspace with Gemini is explicitly HIPAA-eligible. Google's HIPAA Included Functionality list covers Gemini in Workspace. Organizations that sign Google's Business Associate Amendment through the Admin Console can use these services with protected health information. This is currently the most straightforward consumer-tier HIPAA pathway.

Side-by-Side Comparison

HIPAA/BAA Comparison

Beyond the Big Three

This page covers three platforms because they are the ones with mature clinical-privacy pathways and the longest safety track records. But two things outside them now generate more hallway and exam-room questions than anything in the tables above: Grok, and the Chinese AI labs. Neither changes the practical recommendation on this page. Both are worth understanding—for different reasons.

Grok: The AI Your Patients Are Using

Grok is the model built by xAI—folded into SpaceX in February 2026—and it lives where its users live: inside X, in a standalone app, and behind an API. The current flagship is Grok 4.6 (August 12, 2026). Notably, xAI has no healthcare product and no medical vertical; its named enterprise markets are customer support, legal, security, and government. The medical story is entirely organic—and promoted from the very top. Elon Musk has repeatedly told X’s users to upload their medical imaging: “you can upload your X-rays or MRI images to Grok and it will give you a medical diagnosis” in late 2024, and again in February 2026—“just take a picture of your medical data or upload the file to get a second opinion from Grok.” [Forbes]

Two facts to hold at once. First: the popularity is real, but it is your patients, not your colleagues. Roughly one in six US adults now asks a chatbot health questions at least monthly—yet no physician survey measures any Grok adoption at all. The AMA’s 2026 survey found 81% of physicians using AI professionally and named exactly one tool (OpenEvidence); Doximity’s and Sermo’s physician surveys never mention Grok. Second: the capability picture whiplashes with the task. On a public USMLE Step 1 question set, Grok scored best of the five models tested (91.6%)—the study its fans cite. On fifty expert-level radiology spot diagnoses—the exact use being promoted—Grok-4 answered 12% correctly against radiologists’ 83%, near the bottom of the frontier pack. [Preprint] Physician tests of real uploaded scans have found it mistaking tuberculosis for a herniated disk. [Fortune]

The PHI posture is the part to be able to recite. Consumer Grok trains on your conversations by default (opt-out in settings), and xAI’s own privacy policy asks users not to submit health information at all. In August 2025, roughly 370,000 shared Grok conversations—medical questions included—were indexed by Google because share links carried no search-engine protection; shareable links remain a promoted feature. Asked directly on X, Grok itself answered that it is “not a medical professional or HIPAA compliant.” A BAA path does exist—but it is API-only, by application, and requires xAI’s zero-data-retention tier. It does not cover the Grok app, grok.com, or Grok inside X, which are precisely the surfaces patients are being pointed at. So the takeaway matches When Patients Bring AI to the Exam Room: expect Grok readings of labs and imaging to walk in the door, engage with them the way you would any patient research, and keep PHI out of it on your side of the encounter.

The Chinese Labs: Convergence, Price, and Open Weights

The other question—should I care about DeepSeek, Kimi, GLM, Qwen?—got louder in July 2026, when Moonshot’s Kimi K3 release knocked roughly a percent off the Nasdaq in a day and had commentators declaring a second DeepSeek moment. [Fortune] The convergence is real: on Artificial Analysis’s independent index, K3 scores 60 against Claude Opus 5’s 63, Fable 5’s 62, and GPT-5.6 Sol’s 61—the closest an open-weight model has ever sat to the closed frontier—and in nearly every month of 2026 the best open-weight model in the world has been Chinese. [Hugging Face] “China has taken the lead” overstates it for the closed frontier; for open weights it is simply true. Look at who is actually paying, though, and the adoption is price, not preference: on one major developer gateway, DeepSeek’s share of tokens jumped from under 1% to 17% in a single month while its share of revenue stayed near 1%—teams route cheap, low-stakes work to it and keep the frontier models for what matters. [Rest of World]

For clinical use, six facts do most of the work:

• The hosted APIs send your data to China. DeepSeek’s own privacy policy says it plainly: data is collected, processed, and stored “in People’s Republic of China.” No Chinese lab offers a BAA. Patient information into any of these apps or APIs is an impermissible disclosure, full stop. [Policy]
• The model is not the API. The same open weights are served US-hosted by American clouds—Amazon Bedrock, Azure AI Foundry, Google Vertex—where the BAA is with the cloud provider and data stays in the US cloud. A platform BAA makes a service HIPAA-eligible, not a deployment compliant, and model catalogs churn—verify the specific pairing before building on one.
• Open does not mean runnable. K3 is 2.8 trillion parameters of multi-node datacenter hardware. The realistic self-host tier is Qwen3.8-27B (plain Apache 2.0) or DeepSeek V4-Flash—see Running AI Models Locally.
• Open no longer reliably means MIT. Kimi K3 and Qwen’s flagship carry bespoke licenses with revenue-triggered terms; GLM-5.2 is genuinely MIT. Read the license that shipped.
• The medical evidence is China-centric and a generation old. DeepSeek-R1 scored 96% on the Chinese National Medical Licensing Exam—in Chinese, against GPT-4o’s 75%—a language-and-training asymmetry that cuts both ways. [JMIR Med Educ] No physician-rubric evaluation of the current generation exists in either direction.
• No enacted rule restricts your use of them. The US bans are public-sector device and procurement rules (Navy, NASA, Commerce bureaus; Texas, New York, Virginia state devices). Zhipu’s Entity List placement restricts US exports to Zhipu—it does not prohibit Americans from using GLM models. Bills that would go further have not passed.

And the adoption reality-check: no documented US health system runs any of these models, while DeepSeek was deployed locally in roughly 90 Chinese tertiary hospitals by mid-2025. American clinicians are reading about these models, not practicing with them. The reasons to watch anyway are the two that compound: price pressure on the frontier vendors you do use, and a self-host trajectory that makes on-prem clinical AI more plausible every quarter. [Rest of World]

Just Pick One and Start

Here's the honest truth: for most clinical use cases, all three models are capable enough. The differences between them matter at the margins—and those margins shift with every model update anyway.

Don't overthink the choice. Pick based on:

• What you already have access to. Already in Google Workspace? Try Gemini. Have a ChatGPT account from when everyone was talking about it? Use that.
• What your colleagues use. There's value in being able to share prompts and tips with people doing similar work.
• Whichever interface you prefer. Seriously—if one feels better to use, that matters.

You can always switch later. You can use multiple models for different tasks. The skill you develop—learning to prompt effectively, knowing when to trust outputs, building useful workflows—transfers across all of them.

The next section gives you a practical plan for getting started. After that, we'll cover how to evaluate models with your own use cases over time.

Practical Guidance: Your First Month

Week 1: Establish a Baseline

Choose whichever model you have easiest access to. For your first week, use it for low-stakes tasks:

• Summarizing an article you're reading anyway
• Drafting an email you'll heavily edit
• Explaining a concept to yourself before explaining it to a patient
• Brainstorming questions for a meeting

Don't worry about optimization. Just get comfortable with the interaction pattern.

Week 2: Apply Prompting Principles

Revisit the prompting framework from Module 3 and apply it deliberately:

• Give the model a role ("You are helping me prepare for a difficult conversation with a parent")
• Provide context ("The patient is 8 years old with a new ADHD diagnosis")
• Be specific about format ("I need three main points, in language a non-medical parent would understand")
• Include constraints ("Avoid medical jargon; emphasize that this is manageable")

Notice how the quality of outputs changes as you prompt more skillfully.

Week 3: Try Something Harder

Push into a task that actually matters:

• Draft a real (de-identified) clinical note
• Summarize a complex patient case for a referral
• Create patient education materials for a condition you see frequently
• Analyze a clinical guideline and identify key practice implications

Evaluate the output critically. What did it get right? Where did it need correction? What would you prompt differently next time?

Week 4: Compare

Now try a second model with a task you've done before. Use the same prompt and compare outputs. You'll develop intuition for the differences—and often find that your preference is less about the model and more about how you've learned to work with it.

Evaluate With Your Own Use Cases

Here's a truth that benchmark tables and feature comparisons can't capture: the only evaluation that matters is how a model performs on your actual work.

Build Your Personal Test Set

Create a small set of 3-5 tasks that represent your real work:

• A type of document you frequently draft (note, letter, summary)
• A question you commonly need to research or explain
• A complex case or scenario you've worked through before
• Something where you know what "good" looks like

Run these same prompts through different models. Compare the outputs. Which required less editing? Which understood your intent better? Which produced something you'd actually use?

Re-Evaluate After Updates

This is crucial and often overlooked: a model that didn't work for you six months ago might be excellent now. And vice versa—a model you loved might change in ways that don't suit your workflow.

Each major model update (GPT-4 to GPT-4o, Claude 3 to Claude 4, etc.) can significantly change how the model handles specific tasks. Some examples:

• A clinical note format that one model version struggled with might work perfectly in the next
• A type of analysis that frustrated you might work smoothly after an upgrade
• Conversely, a workflow you'd perfected might break when the model changes

When you see announcements about major model updates, revisit your test set. Don't assume your current choice is still the best choice—or that a model you dismissed is still inadequate.

The Cost Question

Let's be direct about money.

Free tiers are sufficient for exploration and light use. If you're using AI occasionally—a few times a week for non-critical tasks—you may never need to pay.

$20/month is the standard paid tier. ChatGPT Plus, Claude Pro, and Google AI Pro all cluster around this price point. At this tier, you get higher usage limits, access to the best models, and priority features. For a professional tool you might use daily, $20/month is modest—less than many software subscriptions with narrower utility. This is where most regular users land.

Premium tiers ($200-250/month) are for power users. If you're running into limits on the $20 tier, pushing complex coding projects, or need maximum model capabilities for professional work, the premium tiers exist. But most users won't need them.

Understanding AI Benchmarks

You'll often see AI companies touting benchmark scores when announcing new models. Headlines declare one model "beats" another on some test. But what do these numbers actually mean—and more importantly, what don't they mean?

What Benchmarks Measure

Benchmarks are standardized tests designed to evaluate specific AI capabilities. They provide a common yardstick for comparing models on tasks like:

• Knowledge recall: Can the model answer questions across academic domains?
• Reasoning: Can it work through multi-step logic problems?
• Coding: Can it write functional code that solves programming challenges?
• Domain expertise: How does it perform on professional exams (medical, legal, etc.)?

What Benchmarks Don't Measure

Here's what no benchmark captures—and what often matters most for real-world use:

• Writing quality: Does the output sound natural and require minimal editing?
• Appropriate uncertainty: Does the model say "I don't know" when it should?
• Instruction following: Does it do what you actually asked, or what it thinks you asked?
• Consistency: Does it give similar quality outputs across multiple attempts?
• Your specific use case: A model that aces coding benchmarks might produce worse clinical notes than one that scores lower.

How to Use Benchmark Data

1. As a rough filter: If a model scores dramatically lower on everything, it's probably less capable overall.
2. For specific tasks: If you primarily need coding help, look at coding benchmarks. For medical questions, look at MedQA scores.
3. With skepticism: Companies often cherry-pick favorable benchmarks. Independent testing (like LMArena) provides more balanced pictures.
4. As a starting point: Use benchmarks to decide which 2-3 models to try, then let your own experience guide your choice.

Key Benchmarks Explained

Benchmark Results (A Snapshot in Time)

Below are scores for flagship models as of April 2026 (per pricepertoken.com and vendor benchmark cards). These numbers will be outdated quickly—new model versions release every few months, and the leaderboard constantly shifts. This table is a good example: it still lists Claude Opus 4.8 and GPT-5.5, both of which have since been superseded (see the July update below). We include it to illustrate the general landscape, not to make a definitive ranking. (pricepertoken.com is a price aggregator—handy for a rough MedQA snapshot, but not an authoritative medical benchmark.) For how LLMs are actually evaluated in medicine—and how rarely they're tested on real patient data—see the JAMA systematic review "Testing and Evaluation of Health Care Applications of Large Language Models" (Bedi et al., 2024).

Common Pitfalls and How to Avoid Them

Quick Reference: Getting Started

Action Items

Before moving to the next module, complete at least one of these:

1. Create accounts on all three platforms (if you haven't already). Even if you end up preferring one, having tried the others gives you useful context.
2. Run the same prompt through all three and compare outputs. Notice the differences in tone, organization, and approach.
3. Try something you actually need: Draft a real email, summarize a real article, prepare for a real conversation. Experience how the tool performs on your actual work.
4. Hit a limit and recover: Deliberately work on something complex enough that your first output isn't good. Practice iterative refinement until you're satisfied.
5. If you're in a healthcare organization: Identify which HIPAA pathway makes sense for your situation and discuss it with appropriate stakeholders.

Summary

ChatGPT, Claude, and Gemini are all capable foundation models that can meaningfully assist clinical work. Each has different ecosystem integrations, pricing structures, and HIPAA pathways. ChatGPT has the largest user base and most extensive ecosystem. Gemini integrates deeply with Google Workspace and offers the most straightforward consumer HIPAA pathway. Claude offers the largest context window outside of Gemini and strong developer tools.

But here's what matters most: the choice between them matters far less than developing skill with whichever you choose. These are tools that respond to how you use them. A well-crafted prompt to any of these models will outperform a vague prompt to the "best" model.

And remember: your evaluation shouldn't be a one-time event. Models improve, workflows change, and what didn't work last year might work beautifully now. Build the habit of periodic re-evaluation with your own real-world test cases.