Claude 3 Opus

Claude 3 Opus arrived on 4 March 2024 at the top of a three-model family alongside Sonnet and Haiku, and it mattered for a simple reason: it was the first model from any lab to beat GPT-4 across the standard benchmark suite. Anthropic reported 86.8% on MMLU with five-shot chain of thought, 50.4% on GPQA Diamond, 95.0% on GSM8K and 84.9% on HumanEval. After a year in which OpenAI's lead had looked structural, a competitor had caught up. The family also introduced the tiering that the whole industry now copies: one expensive and capable model, one balanced, one cheap and fast, all sharing the same context window and interface so that an application can move between them. Opus offered a 200,000-token window, with a million tokens available to selected customers, and near-perfect recall in needle-in-a-haystack retrieval tests. Its most-quoted moment was an accident of that testing. During an internal evaluation, Opus was asked to find a sentence about pizza toppings hidden in unrelated documents, found it, and added that the sentence seemed out of place and was probably inserted as a test of its attention. The exchange was widely reported as a sign of self-awareness; Anthropic's own researcher who published it described it more soberly as a lesson that artificial evaluations were becoming too easy to spot. Opus cost 15 dollars per million input tokens and 75 per million output — expensive even for its moment, and the reason Claude 3.5 Sonnet's arrival three months later was such a shock. Anthropic deprecated the model on 30 June 2025 and retired it on 5 January 2026; requests now fail and the recommended replacement is the current Opus line.

▸Videos

No videos yet.

▸Reviews