Claude Model Family Explained: How to Choose Between Haiku, Sonnet, Opus, and the Mysterious Mythos

Claude Isn't a Single Model — It's an Entire Family

Many people assume Claude is a single AI, but it's actually a full model family built by Anthropic — each member with its own size, speed, and purpose, covering everything from lightweight tasks to flagship-level reasoning. The problem is that most users either pick the wrong model for their needs, or don't even know the other options exist.

A quick note on origins. Claude comes from Anthropic, a company founded by researchers who left OpenAI because they felt AI was moving too fast, too recklessly. They started fresh with a core obsession: making AI genuinely safe, not just powerful. The name "Claude" is a nod to Claude Shannon, the mathematician and father of information theory whose work laid the theoretical groundwork for how computers process language.

Under the hood, Claude works like any large language model — predicting the next token, one word at a time. Its ability to write code, perform analysis, and reason comes from doing that one thing extraordinarily well. What sets it apart is how it was trained: Anthropic developed an approach called Constitutional AI, which instead of relying on endless rounds of human correction, gives the model an explicit set of ethical principles to reason against during training.

Constitutional AI is an alignment training method introduced by Anthropic in 2022. The core idea is to guide a model's self-critique and revision using a written set of principles — the "constitution" — rather than relying entirely on human annotators to score every output. The process works in two stages: first, the model evaluates and rewrites its own responses according to the constitutional principles (a supervised learning phase); then a variant of reinforcement learning from human feedback (RLHF) is used, where the model learns to choose the better response among candidates based on those principles, generating preference data. The benefit is a dramatic reduction in dependence on human labeling, while making safety constraints more transparent — the principles are publicly available, rather than buried in the subjective judgments of thousands of annotators. This is the fundamental reason Claude sometimes comes across as more "principled" or "cautious" than other models.

Haiku: The Lightweight Built for Speed

Haiku is the smallest, lightest, and fastest model in the family — and the one most users forget exists. It's built for speed, bulk processing, and tasks where you need an answer right now.

Practical use cases include: quick summaries, extracting key information from documents, answering straightforward questions, running batch processing tasks, and handling any repetitive work you'd rather not do manually. A fitting analogy: Haiku is like an intern who handles your emails, scheduling, and copy-pasting — solid at the day-to-day stuff, but don't expect them to run the meeting. Its weakness is depth: throw a tricky engineering problem at it or ask for nuanced analysis, and it will visibly struggle.

Sonnet: The Reliable Everyday Workhorse

Sonnet is the community's consensus "default choice." The current version, Sonnet 4.6, was released in February 2026, and features a context window of roughly one million tokens — equivalent to approximately 2,000 to 3,000 pages of text in a single conversation.

Its range of applications is broad: developers use it for senior-engineer-level coding tasks, writers use it to produce substantial content, researchers use it to analyze entire documents, and businesses run it as an operational brain. The community consensus is pretty clear — "use Sonnet for 90% of things; only switch to Opus when you're making decisions that are hard to reverse."

One thing worth noting about Sonnet's personality: it won't flatter you. It won't tell you your half-baked ideas are brilliant — it'll point out what's wrong. If you've never used Claude before, starting with Sonnet is the safest bet. It hits the sweet spot between intelligence and speed.

A context window refers to the maximum number of tokens a model can process simultaneously in a single conversation. A one-million-token context window means you can feed an entire book, a full codebase, or dozens of reports into the model at once without worrying about it "forgetting" earlier content. Tokens are not the same as words — in English, one word typically maps to one or two tokens; in Chinese, each character maps to roughly one to two tokens. So one million tokens corresponds to roughly 700,000–800,000 English words, or about 400,000–500,000 Chinese characters. The practical value of an ultra-long context window is the ability to perform cross-document analysis, full-codebase review, or coherent long-form content creation in a single request — dramatically reducing the risk of information loss from truncated context.

Opus: Flagship Performance, and the Most Expensive

Opus is the flagship model that most users have never actually experienced — not because they don't want to, but because it's priced accordingly. Opus 4.7 was released in April 2026, with the same one-million-token context window, at roughly $5 per million input tokens.

The performance numbers are impressive: on SWE-Bench — which measures whether an AI can fix real GitHub bugs the way an actual software engineer would — Opus 4.7 scores 87.6%. On GPQA Diamond, a set of questions challenging enough to stump PhD-level scientists, it scores 94.2%. Engineers have had it run for seven straight hours refactoring entire open-source codebases without losing focus. When Sonnet starts struggling, switching to Opus often just solves the problem outright.

The trade-offs are real: Opus is expensive, burns through tokens fast, and on the free tier you might hit your rate limit after just a few prompts. Developers also half-jokingly note that Opus will write three paragraphs explaining why it's being careful before it actually answers your question. But when you're dealing with something genuinely hard, nothing else currently comes close.

SWE-Bench is a benchmark released by a Princeton University research team to measure AI models' ability to solve real-world software engineering tasks. It draws thousands of real Issues and corresponding code fix patches from GitHub, and requires the model to locate the problem, modify the code, and pass the original test suite — without any human prompting. The reason this benchmark is considered highly credible is that it tests end-to-end engineering execution, not multiple choice or Q&A: the model must understand the structure of an unfamiliar codebase, grasp the bug's context, and generate a working fix. GPQA Diamond, by contrast, is a set of multiple-choice questions at doctoral-level difficulty in biology, chemistry, and physics — even domain experts average only around 65% correct, which makes a score of 94.2% particularly striking.

Mythos: The "World-Ender" That's Locked Away

The most mysterious member of the family is Mythos — it's not available to the public, and only a small number of partners can access it through a restricted internal program.

According to available reporting, Mythos's reasoning and software engineering capabilities are considered too powerful to release safely. It reportedly scanned all major operating systems and browsers and discovered thousands of critical vulnerabilities that had gone undetected for decades. On "Humanity's Last Exam" — described as the hardest knowledge benchmark ever created — it scored 64.7%, far ahead of any publicly available model. In early internal testing, its ability to read code, find vulnerabilities, and compromise systems reportedly exceeded that of top human hackers. For that reason, it won't be released publicly until Anthropic has established sufficiently robust safety controls.

Note: Information about Mythos comes from a single video source and lacks corroboration from public documentation. Readers should treat these claims with appropriate skepticism.

Claude's Weaknesses and How to Pick the Right Model

Claude isn't without its flaws. It can be frustratingly slow, and the free tier may cut you off after just two or three prompts — something that regular ChatGPT or Gemini users rarely run into. It only handles text and code; there's no image or video generation. It can also be overly cautious — push the boundaries even slightly and it will hesitate more than ChatGPT or Gemini, even for entirely legitimate requests. If you want blunt, no-nonsense answers without the hedging, Grok's distinctive personality might suit you better.

Here's a clear decision framework for picking the right model:

• Quick tasks, batch processing, repetitive work → Use Haiku — fast and cheap
• The vast majority of everything else → Use Sonnet — the best balance of intelligence and speed
• Tricky engineering, financial modeling, hard-to-reverse decisions → Pay up for Opus
• Mythos → Not an option — it's unavailable to general users for now

Once you understand the family structure, you'll stop thinking of Claude as one generic "AI" and start deploying the right member for each job.