Anthropic made Claude 2.1 available on November 21, 2023 via its API and the claude.ai chat interface. The update targets enterprise use-cases and introduces several major changes: an industry-leading 200,000 token context window, significant reductions in model hallucinations, and a new beta feature for tool use. The company also noted updates to pricing intended to improve cost efficiency for customers, without providing detailed tariff tables in the announcement.
200K-token context window
A central feature of Claude 2.1 is the 200,000 token context window, which Anthropic estimates corresponds to roughly 150,000 words or more than 500 pages of material. This expansion responded to user requests to work with much larger documents—examples cited include entire codebases, SEC S-1 filings, or long literary works.
Processing messages of that length is technically complex and, according to Anthropic, represents an industry first. They caution that tasks that would typically take humans hours to complete may still take Claude a few minutes, and expect latency to fall as the technology matures.
Reduced hallucination and improved accuracy
Claude 2.1 shows measurable improvements in honesty and accuracy: Anthropic reports a 2x reduction in false statements relative to Claude 2.0. The company evaluated honesty with a large set of complex factual questions designed to probe known model weaknesses and used a rubric that distinguishes incorrect claims from admissions of uncertainty. Claude 2.1 was substantially more likely to decline to answer or to express uncertainty than to provide incorrect facts.
The model also improved comprehension and summarization for long, complex documents. In Anthropic’s evaluations, Claude 2.1 produced 30% fewer incorrect answers and was 3–4x less likely to mistakenly conclude that a document supported a particular claim. The company emphasized that, while these gains are encouraging, increasing precision and reliability remains a top priority.
API tool use (beta)
Anthropic introduced a beta tool use feature that enables Claude to interact with developer-defined functions, APIs, web sources, and private knowledge bases. The aim is to improve interoperability so Claude can fit into existing processes and products.
Use-case examples include:
- using a calculator for complex numerical reasoning,
- translating natural language requests into structured API calls,
- answering questions by searching databases or invoking a web search API,
- taking simple actions in software through private APIs,
- connecting to product datasets to make recommendations and assist purchases.
Tool use is in early development; Anthropic is building developer features and prompting guidelines and requests user feedback to refine the capability.
Developer experience: Workbench and system prompts
Anthropic has also refined the developer Console and introduced Workbench, a playground-style product to iterate on prompts more quickly. Workbench lets developers create multiple prompts, switch among them for different projects, and automatically saves revisions to preserve history. Developers can generate code snippets to deploy prompts via Anthropic’s SDKs.
The company additionally introduced system prompts, which allow users to provide persistent, custom instructions to Claude. System prompts help set context, define personalities or roles, and produce more consistent, structured outputs aligned with user needs.
Availability and limitations
Claude 2.1 is available now via Anthropic’s API and at claude.ai for both free and Pro tiers. The 200K-token context window is reserved for Claude Pro users, who can upload larger files than before. Anthropic says it looks forward to the enterprise and research use-cases these features will enable, while continuing to prioritize safety and technical refinement.
