Claude is a family of state-of-the-art large language models developed by Anthropic, built using Constitutional AI (CAI)—a safety-first training approach that embeds ethical principles directly into the model during both pre-training and fine-tuning. Unlike traditional RLHF-only methods, Constitutional AI uses AI feedback to self-critique and revise responses before human evaluation, creating models that are helpful, harmless, and honest by design. Claude excels at complex reasoning, coding, multimodal understanding, and agentic workflows, offering context windows up to 1 million tokens (Opus 4.7, Sonnet 4.6) with specialized features like adaptive thinking, prompt caching, tool use, and the Claude Managed Agents platform for production-grade autonomous agents. The current flagship, Claude Opus 4.7 (April 2026), introduces adaptive-only thinking, high-resolution vision (3.75MP), a xhigh effort level, and breaking API changes—while the research-preview Claude Mythos Preview demonstrates frontier cybersecurity capabilities through Project Glasswing.
What This Cheat Sheet Covers
This topic spans 20 focused tables and 132 indexed concepts, 112 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.
A jump-to index of every table row in this cheat sheet.
An interactive map of every table and concept in this topic.
Table 1: Model Variants
Picking the right Claude model is the first decision in any project — and it's mostly a trade between raw capability, speed, and cost. Opus is the heavyweight for hard reasoning and coding, Sonnet is the everyday workhorse that balances both, and Haiku is the cheap, fast option for high-volume tasks. The exact model strings, context windows, and per-token prices below are what you'll actually pass to the API.
| Model | Example | Description | |
|---|---|---|---|
claude-opus-4-7 | • Most capable generally available model (April 16, 2026) • adaptive thinking only; high-res vision (3.75MP / 2576px); 128K max output • 1M context window at standard pricing • 70% CursorBench (vs 58% Opus 4.6); 3× more production tasks solved • $5/$25 per MTok; updated tokenizer uses 1.0–1.35× more tokens. | ||
claude-sonnet-4-6 | • Best balance of speed and intelligence; default for most production workloads • extended thinking + adaptive thinking; 1M context; 64K max output • $3/$15 per MTok; ~30–50% faster than Sonnet 4.5. | ||
claude-haiku-4-5 | • Fastest and most cost-effective model • extended thinking + adaptive thinking; 200K context; 64K max output • $1/$5 per MTok; best for high-throughput tasks, classification, customer support. | ||
claude-opus-4-6 | • Previous flagship; still supported but recommend migrating to Opus 4.7 • extended thinking + adaptive thinking; 1M context; 128K max output • $5/$25 per MTok; 76% needle-in-haystack at 1M tokens. | ||
Invitation-only (Project Glasswing partners) | • Anthropic's most capable frontier model—not publicly available • exceptional at autonomous vulnerability discovery across major OSes and browsers • used by AWS, Google, Microsoft, CrowdStrike, etc. for defensive cybersecurity • $25/$125 per MTok for Glasswing participants after initial credits. |