Claude 2

// Overview

Claude 2 is the second-generation model in Anthropic's Claude (legacy 1–2.x) line, released on 2023-07-11 as the first Claude available directly to the public — through the claude.ai chat (beta in the US and UK) and a paid API. It was a text-in, text-out model; the claude.ai interface added file uploads (PDF, TXT, CSV) so users could ask questions over documents, but the model itself had no image, audio, or video input.

Its headline feature was a 100,000-token context window (Anthropic trained it to handle up to 200K tokens but shipped 100K at launch), enough to load hundreds of pages — roughly 75,000 words — into a single prompt. Claude 2 also posted large jumps over Claude 1.3 on Anthropic's evaluations: 71.2% on the Codex HumanEval Python coding test (up from 56.0%), 88.0% on the GSM8K grade-school math set (up from 85.2%), and 76.5% on the multiple-choice Multistate Bar Exam. On the GRE it scored in roughly the 95th percentile on Verbal and 91st on Analytical Writing. Its training data cuts off in early 2023.

Claude 2 was a paid API-only and chat model — Anthropic never released its weights or disclosed its parameter count or architecture beyond it being transformer-based. Anthropic deprecated it (API id claude-2.0) on 2025-01-21 and retired it on 2025-07-21, alongside Claude 2.1 and Claude 3 Sonnet; the model is no longer callable and Anthropic recommends migrating to current Claude models.

// Benchmarks

1. MMLU (5-shot CoT)78.5%
2. Codex HumanEval (Python, 0-shot)71.2%
3. GSM8K (grade-school math, 0-shot CoT)88%
4. ARC-Challenge (5-shot)91%
5. TriviaQA (5-shot)87.5%
6. Multistate Bar Exam — MBE (5-shot)76.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Standard claude-2.0 API pricing (Anthropic, Nov 2023). Model retired 2025-07-21 — no longer billable.

Pricing source ↗

// Strengths

• 100K-token context window — read hundreds of pages, long documents, or short books in one prompt (a major jump over the prior 9K window)
• Strong-for-its-era coding: 71.2% on Codex HumanEval, up from 56.0% on Claude 1.3
• Solid quantitative and verbal reasoning: 88.0% GSM8K math and 76.5% on the multiple-choice bar exam
• First publicly available Claude — usable via claude.ai chat with file uploads (PDF/TXT/CSV) as well as the API
• Anthropic reported it was about 2x better than Claude 1.3 at giving harmless responses

// Best for

• Summarizing and Q&A over long documents — technical manuals, financial reports, contracts, or short books
• Code generation, explanation, and review (e.g. early Sourcegraph Cody used Claude 2's long context)
• Drafting and editing long-form text: memos, letters, and stories up to a few thousand tokens
• Knowledge-work assistants needing more context than the ~9K-token models of the time could hold

// How to access

// Claude (legacy 1–2.x) — every version

The full lineage of the Claude (legacy 1–2.x) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.