Anthropic Launches Claude 2, With 100,000-Token Context Window

On July 11, 2023, Anthropic introduced Claude 2; the release offers a useful lesson in measuring a context window without mistaking input capacity for reliable understanding.

Anthropic launched Claude 2 today, the new version of its conversational assistant, giving the public direct access through its own website in the United States and the United Kingdom. Until now, most people who encountered Claude did so indirectly, through third-party products like Slack or Notion, or via paid API access aimed at businesses. With today's launch, anyone in those two countries can sign up and talk to the model directly.

A much longer memory

The number that defines this launch is 100,000 tokens: the amount of text Claude 2 can hold in mind at once within a single conversation. A token roughly corresponds to a chunk of three or four characters, so 100,000 tokens works out to around 75,000 words — about the length of a mid-sized novel. Primary source

That context window is what lets users, for instance, paste in a lengthy document, a technical report, or even several chapters of a book, and ask the model to summarize, analyze or answer questions about all of it without losing the thread. For comparison: GPT-4, OpenAI's flagship model, offered an 8,000-token window in its standard version, expandable to 32,000 in its pricier variant. Claude 2 multiplies that figure several times over, making it the model with the largest publicly available context window to date. Primary source

Gains in coding, math and reasoning

Anthropic has pointed to concrete improvements over Claude 1.3, its previous version. On the Codex HumanEval coding benchmark, which measures the ability to write functional Python code from a description, Claude 2 jumped from 56% accuracy to 71.2%. On the multiple-choice section of the US bar exam, it rose from 73% to 76.5%. On the GRE, the standardized test used for graduate school admissions in the US, the most striking leap came in quantitative reasoning: from the 25th percentile to the 95th percentile. Primary source

These figures come directly from the company alongside the announcement, and they serve as a gauge of where the model is headed: less toward casual conversation, and more toward analysis, problem-solving and code generation — tasks that until now had been more the domain of models like GPT-4.

Who's behind Claude

Anthropic was founded in 2021 by Dario Amodei and Daniela Amodei, both former research leaders at OpenAI, along with another group of researchers who left the same company. From the start, the company has put safety at the center of its messaging: its model is trained using a technique it calls Constitutional AI, a method by which the system itself evaluates and corrects its own responses according to a set of explicit principles, rather than relying solely on humans to review and label every problematic answer.

Google invested roughly $300 million in Anthropic in 2022, and the company has continued raising capital from other investors since then. That financial backing has allowed it to compete in the large language model race against giants like OpenAI (backed by Microsoft) and Google, despite being a much younger and smaller company than either. Primary source

What changes for users

Until now, most people who had tried Claude did so indirectly — through Slack, the coding tool Sourcegraph, or other products that integrated the model via API. With today's launch, Anthropic starts competing directly with ChatGPT on the very ground where it has dominated for months: open, direct public access to a conversational assistant of its own.

For now, availability limited to the US and UK marks the real scope of the announcement. For other markets, including Spain, Claude 2 will for now continue to reach users only through businesses and developers integrating its API, not as a direct-to-consumer product.

A large window does not guarantee that the model finds what matters

A context window says how much material fits in a request, not how evenly the system recalls it or how much it can return. The Claude 2 announcement separated those properties: it accepted up to 100,000 input tokens and described outputs of several thousand tokens. That distinction prevents a common inference. Loading an entire book does not guarantee a complete reproduction or a correct answer about every detail.

To test long context, build a document whose contents are known. Place distinctive facts near the beginning, middle and end; add two similar passages that could be confused and one question whose answer is absent. Ask the model to attach the exact supporting passage to every conclusion. This measures retrieval, attribution and abstention, three different capabilities that a polished summary demo often hides.

Order matters too. Repeat the same questions after moving the relevant passages. If the model succeeds when evidence is at the end but misses it when buried in the middle, its practical limit is smaller than the headline number. Record the version, settings and actual prompt length as well: services change, and a result without those coordinates cannot be compared.

How to read benchmark points

HumanEval, professional exams and academic tests provide signals, but they are not interchangeable with everyday work. Ask which version was evaluated, how many attempts it received, which tools it could use and whether the company selected the best run. An internal improvement over Claude 1.3 shows progress under that protocol; it does not establish superiority across every programming language, jurisdiction or document type.

The transferable skill is to translate “more context” into a verifiable retrieval test. Before trusting any model with a contract, investigation or repository, hide known facts in different positions, demand traceable citations and include an unanswerable question. If the system cannot show where a claim came from or decline to answer, window size does not solve the problem.

There is another constraint: material placed in context competes with instructions and the space reserved for the answer. Before loading a case file, divide it into sections, preserve headings and numbering, and first request an index of claims with locations. Then ask focused questions. If the model supplies a quotation that does not exist, the remedy is not to repeat the question more forcefully but to return to the document and narrow the scope.

For sensitive documents, technical evaluation does not replace a data decision. Determine whether the provider retains inputs, for how long, who can access them and whether they are used for training. A large window encourages uploading more information at once, which increases the impact of a misunderstood privacy setting. Minimising the passages sent remains a useful defence.

The same test works for future, larger windows. Increase length while keeping questions and evidence constant, chart accuracy by position and record where useful performance degrades. The headline number then becomes a curve that helps decide how much context is worth sending.