Most people meet Claude through a chat box and assume there is one of it. There isn’t. Anthropic ships Claude as a family of models, each with its own size, price and set of strengths, and which one answers your question depends on where you typed it. The free web app, the Pro plan, the desktop client and the API do not always route to the same brain.
That difference is easy to feel but hard to name. Two people can ask the same question on the same afternoon and get noticeably different quality, purely because one landed on Opus and the other landed on Haiku. Knowing the lineup makes you a sharper user of all of it, and it saves money.
What Anthropic actually ships under the Claude name
Anthropic was founded in 2021 by a group of former OpenAI researchers, including siblings Dario and Daniela Amodei. Its models are trained with a technique the company calls Constitutional AI: the model is given a written set of principles and taught to critique and revise its own drafts against them. In practice that produces a family that hedges more than some rivals and refuses less awkwardly than the earliest GPT versions.
The structure is consistent across generations. You get a generation number (3, 3.5, 3.7, 4) and a tier name: Opus, Sonnet or Haiku. Read the tier first and the number second. A newer Haiku is fast and cheap; an older Opus is still slow and expensive. They are not interchangeable.
Opus, Sonnet, Haiku: what the tiers mean
Claude Opus
The largest tier, and the one Anthropic points at hard reasoning, long multi-step analysis and research work. Opus 4.1 sits around $15 per million input tokens and $75 per million output tokens, which sounds abstract until you push a 200,000-token document through it a few hundred times. You reach for Opus when a wrong answer costs more than the tokens do.
Claude Sonnet
The middle tier and the sensible default for most people. Sonnet balances capability against price (roughly $3 in and $15 out per million tokens) and has quietly become Anthropic’s flagship for software work. Claude 3.5 Sonnet was the release that pulled a lot of developers away from GPT-4, and Sonnet 4.5 pushed the same argument further.
Claude Haiku
The small, fast tier. Haiku handles classification, data extraction, tagging and summarising at scale, where latency matters and the task is well-defined. It costs a fraction of Sonnet and is often good enough. Push it on open-ended writing and you will notice the prose flatten out.
How the generations actually changed
Claude 3 arrived in March 2024 and made a 200,000-token context window the baseline, roughly 150,000 words or a 500-page book in a single prompt. That was the number that got enterprises to take notice, because it turned “summarise this PDF” into “read this entire contract set.”
Claude 3.5 Sonnet landed in mid-2024 with sharper coding and faster responses; 3.5 Haiku followed later that year. Claude 3.7 Sonnet in February 2025 introduced extended thinking, where the model spends visible computation working through a problem before it answers. On debugging and multi-step maths the improvement was obvious rather than marginal.
Claude 4 shipped in May 2025 as Opus 4 and Sonnet 4, with Opus 4.1 in August, Sonnet 4.5 in September and Haiku 4.5 in October. The 4.x line leaned hard into agentic work: running tools in loops, editing files, holding a coherent thread across multi-hour sessions. Sonnet 4.5 also pushed the context window to 1 million tokens in beta.
Context windows, and the catch nobody mentions
A bigger window is not automatically a better answer. Cost scales with input, so a million-token prompt on Sonnet is a real invoice. Models also lose precision in the middle of very long contexts, a pattern documented across every vendor. Stuff a 300-page document in and bury the actual question on page 180, and you should expect mush. Put instructions at the start or the end, and pull out the two relevant sections when you can.
Matching the model to the job
- Drafting, editing and tone work: Sonnet, with Opus for anything that goes out under your name.
- Production code: Sonnet. It writes tests more reliably than it explains them, and it is the tier most developers settle on.
- Bulk classification across thousands of rows: Haiku, batched through the API.
- Long autonomous runs: Sonnet 4.5 or Opus 4.1, both built for tool loops.
- Quick lookups on your phone: whatever the app defaults to, usually Haiku or Sonnet.
Where Claude models genuinely earn their keep
Prose is the strongest suit. Claude tends to produce sentences with real rhythm, and it follows style instructions like “shorter paragraphs, no rhetorical questions” better than most. Our Claude AI chatbot review digs into where that strength shows and where the safety tuning starts sanding off the edges.
Coding is the second. Long file edits, refactors across a dozen files and patient debugging are areas where Sonnet holds up well. The honest limits are covered in Claude AI for coding, including the cases where it confidently invents an API that never existed.
Long-document work is the third. Contracts, research papers, transcripts, entire codebases. That 200K baseline is the reason a lot of legal and research teams started paying Anthropic rather than OpenAI in the first place.
Getting set up, and what happens to your data
The consumer app is the easiest entry point, and the setup quirks are worth knowing before you sign in, particularly around Projects and file uploads. Our walkthrough of the Claude AI app covers that in detail.
Privacy is the other thing to sort out early. Consumer chats and API traffic are handled differently, and the settings that control retention are not all in the same menu. There is a clear breakdown of what you can actually switch off in Claude AI privacy.
Claude against Gemini and Grok
The obvious question is whether any of this beats the alternatives. Gemini wins on integration with Google’s stack and on very long audio and video inputs. Grok wins on real-time information and a looser tone. Claude tends to win on writing quality and on sustained coding sessions, and it loses on anything that needs a live feed of today’s news unless you bolt a search tool onto it. If you’re weighing a switch, Claude vs Gemini breaks the decision down by workflow rather than benchmark scores.
A setup that works from day one
Start on Sonnet and stay there until something feels wrong. If answers come back thin on a well-specified task, drop to Haiku and check whether the extra capability was ever doing the work. If a task keeps needing three follow-up corrections, move up to Opus and see whether the first answer improves. Most people never need to touch the model picker at all, and the ones who do usually need it for a narrow, expensive slice of their work.
Turn on extended thinking for anything with numbers in it. Feed long documents in whole rather than in fragments, but put your actual question last. Keep a running list of prompts that produced something you would actually use, because the gap between a mediocre Claude session and a good one is usually the prompt, not the tier.