Claude 2
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 1,420 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 1,420 words
Add missing citations, update stale details, or suggest a clearer explanation.
Claude 2 is a large language model released by Anthropic on July 11, 2023, the second major version of the company's Claude family of conversational AI systems and the first Claude model offered directly to the general public through a free beta chat interface at claude.ai. [1][2][3] It carried a 100,000 token context window (about 75,000 words, the length of a short novel), scored 71.2% on the Codex HumanEval Python coding test, and reached 76.5% on the multiple-choice section of the US bar exam. [1][4] Compared with the earlier Claude 1 (specifically the Claude 1.3 release), Claude 2 reported gains in coding, mathematics, and reasoning while keeping the 100,000 token context window introduced in the prior generation. [1][4]
Anthropic, founded in 2021 by former OpenAI staff including siblings Dario and Daniela Amodei, had released the first version of Claude in March 2023. That model was available only to a limited set of approved partners and business customers, not to ordinary users. [2][5] Claude 2 followed about four months later. Anthropic described it less as a ground-up rebuild than as a refinement of its predecessor: TechCrunch characterized the system as "a tweaked version of Claude 1.3," the product of roughly two years of work rather than a wholly new architecture. [4]
The release landed in the middle of an intense period of competition among chatbot makers, with OpenAI's GPT-4 and Google's Bard already in the market. Anthropic counted Google among its investors, and press coverage framed Claude 2 as part of the broader contest between OpenAI's backers and Google. [3][6]
Claude 2 accepts prompts of up to 100,000 tokens, the same limit as Claude 1.3. [1][4] Anthropic put the figure plainly in its announcement: "You can now input up to 100K tokens in each prompt, which means that Claude can work over hundreds of pages of technical documentation or even a book." [1] In practical terms that allowed users to submit on the order of 75,000 words at once, roughly the length of a novel such as The Great Gatsby, so the model could read through long documents, technical manuals, or transcripts in a single pass. [1][4] Output was shorter: the model could generate up to about 4,000 tokens, on the order of 3,000 words, in one response. [4] TechCrunch noted that the underlying system could in principle handle around 200,000 tokens, but that this was not enabled at launch. [4]
Anthropic positioned Claude 2 for written tasks including summarization, search, question answering, and coding, and the claude.ai interface accepted uploaded files such as PDFs (up to five files in a single chat) so users could ask questions about their contents. [1][3][7] The model was trained on a mix of public web data, licensed third-party datasets, and data voluntarily supplied by users, with training data extending into early 2023. [4][8]
Anthropic published a set of exam and benchmark results comparing Claude 2 against Claude 1.3. The headline figures were a higher score on the multiple-choice portion of the US bar exam and a large jump on a Python coding test. [1][4] The company also reported standardized-test percentiles on the GRE.
| Evaluation | Claude 2 | Claude 1.3 |
|---|---|---|
| Bar exam, multiple-choice section | 76.5% | 73.0% |
| Codex HumanEval (Python coding) | 71.2% | 56.0% |
| GSM8K (grade-school math) | 88.0% | 85.2% |
| GRE reading and writing | above 90th percentile | not reported |
| GRE quantitative reasoning | near the median applicant | not reported |
Sources: Anthropic's announcement and model card, with the bar exam, HumanEval, and GSM8K comparisons also reported by TechCrunch. [1][4][8]
The coding result drew the most attention, since the move from 56.0% to 71.2% on Codex HumanEval represented the clearest single-metric improvement over the previous generation. [1][4] Several outlets read the combination of these scores and the long context window as evidence that Claude 2 had moved into the same competitive tier as OpenAI's GPT-3.5 and GPT-4 systems. [6]
Before Claude 2, access to Claude had been gated to selected businesses and developers. With this release Anthropic opened a public beta chat experience at claude.ai, free to use, initially restricted to users in the United States and the United Kingdom, with wider availability promised later. [1][3] This was the debut of the consumer-facing claude.ai web product that the company continued to develop in later years.
Developers could also reach Claude 2 through Anthropic's API. The company said API pricing was unchanged from Claude 1.3; TechCrunch put the rate at roughly 4.65 cents to generate about 1,000 words. [1][4] A faster, lower-cost sibling model, Claude Instant, remained available for lighter workloads alongside the larger Claude 2. The free public beta predated Anthropic's paid consumer subscription, Claude Pro, which the company introduced in September 2023. [9]
Anthropic trained Claude 2 with the same broad safety approach used across the Claude line, combining reinforcement learning from human feedback with the company's Constitutional AI method, in which the model critiques and revises its own outputs against a written set of principles rather than relying solely on human-labeled examples of harmful content. [5][10] The technique was one of the features that Anthropic and the press used to distinguish Claude from competing systems. [6]
On safety metrics, Anthropic said internal red-teaming found Claude 2 was "2x better at giving harmless responses" than Claude 1.3. [1][4] The company was candid that the model was not infallible: like other large language models it could still produce confabulations, reflect bias, make factual errors, and be coaxed past its guardrails. [5]
Coverage generally treated Claude 2 as a credible challenger to ChatGPT rather than a clear leader. CNBC noted that Anthropic was opening its technology to consumers for the first time, shortly after raising substantial funding. [3] TIME and others walked readers through how to use the new chatbot and how its capabilities compared with GPT-4 and Bard, with particular interest in the 100,000 token context window, which exceeded what most rivals then offered to typical users. [6][11] The free price of the beta, against the paid tiers required for the most capable competing models, was a recurring talking point. [6]
Anthropic released an incremental update, Claude 2.1, on November 21, 2023. That version doubled the context window to 200,000 tokens (roughly 150,000 words, or over 500 pages of text), added system prompts and a beta tool-use feature, and reduced hallucination rates: Anthropic reported "a 2x decrease in false statements compared to our previous Claude 2.0 model." [12][13] Access to the full 200,000 token window was reserved for paying Claude Pro subscribers at launch. [13]
The Claude 2 series was superseded in March 2024 by the third generation, led by Claude 3 Opus, which moved the family to a new three-tier lineup (Haiku, Sonnet, Opus) and added image understanding. [2][14] Anthropic said the mid-tier Claude 3 Sonnet was "2x faster than Claude 2 and Claude 2.1" with higher intelligence, while Claude 3 Opus showed "a twofold improvement in accuracy" over Claude 2.1 on challenging open-ended questions. [14] Later Anthropic generations, including Claude 3.5, Claude 4, and the Opus 4.x line, continued from that lineup.