A focused timeline of the model family most of this handbook is addressed to.
Anthropic was founded in mid-2021 by Dario and Daniela Amodei, Tom Brown, Jared Kaplan, Sam McCandlish, Chris Olah, Jack Clark, and others who had led much of the GPT-3 effort at OpenAI. Their founding thesis: frontier AI must be built by people doing serious safety research — because if powerful systems are coming regardless, you want the builders and the safety researchers to be the same people.
Claude is the outward-facing product of that thesis. This chapter is a tight history, ending at the model probably generating the pages in front of you.
In plain English. Where OpenAI's public story is "build it and ship it," Anthropic's story is "build it carefully, and publish what you learn about why it's safe." The engineering shows up in both.
mindmap
root((The Claude line))
Safety
Constitutional AI
Responsible Scaling Policy
Interpretability research
Refusal calibration
Capability
Long context (100k -> 1M+)
Multimodality
Reasoning / extended thinking
Coding leadership
Product
Claude.ai chat
API + SDKs
Claude Code CLI
Claude in Chrome
Claude in Excel
Cowork desktop
Agent SDK
timeline
title The Claude family
2023 Mar : Claude 1 - Constitutional AI
2023 Jul : Claude 2 - 100k context
2023 Nov : Claude 2.1 - 200k, lower hallucinations
2024 Mar : Claude 3 (Haiku/Sonnet/Opus)
2024 Jun : Claude 3.5 Sonnet + Artifacts
2024 Oct : 3.5 Sonnet new + 3.5 Haiku + Computer Use
2025 Feb : Claude 3.7 - Extended thinking
2025 May : Claude 4 family (Opus 4, Sonnet 4)
2025 Sep : Claude 4.5 Sonnet / Haiku
2025 Q4 : Opus 4.6
2026 Q2 : Opus 4.7 (today)
Claude 1 shipped with 9k of context and an API-only interface. Its distinguishing trick was Constitutional AI (CAI): instead of teaching the model to be helpful and safe only through human feedback, Anthropic wrote a constitution of principles and trained the model to critique and revise its own outputs against those principles.
flowchart TB
A[SFT model] --> G[Generate response]
G --> CR[Self-critique
against constitution]
CR --> R[Revise response]
R --> D[Dataset of revised responses]
D --> F[Fine-tune]
F --> A2[Aligned model]
The payoff was practical: you could adjust behavior by editing prose rules, not by rehiring a thousand raters. Every later Claude model's behavior has been shaped by evolving versions of this constitution.
Claude 2 shipped with a 100,000-token context window — a first at the time, about 10× the frontier competition. For the first time, you could paste an entire book into a chat and ask about it.
This was more than a marketing number. It changed what products were possible:
Claude 2.1 (November 2023) doubled the window again (200k) and materially reduced hallucination rates.
Claude 3 introduced the Haiku / Sonnet / Opus tiered naming that other vendors later copied. Each tier serves a different latency/cost/capability sweet spot:
| Tier | Meant for | Typical use |
|---|---|---|
| Haiku | Cheapest, fastest | High-volume classification, extraction, routing |
| Sonnet | Balanced | Default workhorse for most apps |
| Opus | Most capable, most expensive | Hardest reasoning, agents, frontier work |
Claude 3 also introduced vision — image understanding with text. Combined with long context, it was the first model that felt like it could "read" a PDF the way a human reads a PDF: diagrams, tables, handwriting and all.
flowchart LR
subgraph Claude 3 family
H[Haiku] --> S[Sonnet] --> O[Opus]
end
H -.fast/cheap.-> S
S -.balanced.-> O
O -.max capability.-> end1[agents, hard reasoning]
H -.high volume.-> end2[classification, extraction]
S -.most apps.-> end3[chat, RAG, copilots]
Claude 3.5 Sonnet was, by several benchmarks, the first non-frontier-name ("Opus") model to match or beat the frontier. It also brought two product-level shifts:
In October 2024, Anthropic shipped three things on the same day:
Computer Use was an early, rough capability — slow, brittle on real websites, and easy to mock. It was also the first clear signal that general-purpose autonomous agents driving real UIs were a matter of months, not years.
3.7 Sonnet added extended thinking — the Claude equivalent of OpenAI's o1 pattern. The model could spend thousands of hidden tokens reasoning before answering, configurable per request. This unified two previously separate model families (chat vs reasoning) into one model with a dial.
flowchart LR
Q[Prompt] --> M[Claude 3.7+]
M -->|thinking = 0| A1[Fast chat answer]
M -->|thinking = low| A2[Short CoT]
M -->|thinking = high| A3[Deep reasoning]
Claude 4 — Opus 4 and Sonnet 4 — was the first release specifically framed around long-horizon agentic tasks. Anthropic reported successful autonomous runs of hours, with the model planning, executing, self-correcting, and completing substantial engineering tickets without intervention.
Other notable shifts:
Through 2025, Anthropic shipped a steady cadence:
Five recurring themes:
As of this handbook's writing, Claude is accessible via:
Put together, these are the main ways most of this handbook's "things I can do with an AI" actually get done.
This is an opinion chapter in a mostly-neutral handbook. Honestly:
Competitors are excellent and catching up constantly. The right default in 2027 may be different. In April 2026, Opus 4.7 is the one I'd give a working engineer without caveats.