On Thursday, Antarubor Claude Obus 4 and Claude Sonit 4 released the company’s return to the largest typical versions after focusing mainly on medium -term Sonata variables since June last year. The new models represent what the company calls the most capable coding models, with the OPUS 4 design for long -term complex tasks that can work independently for hours.
Alex Albert, the head of Claude’s relationships in man, told Ars Technica that the company chose to revive the OPUS line because of the increasing demand for AI Agent applications. “In all companies that build things, there is a really large wave of these agents that appear and high demand for intelligence,” Albert said. “I think OPUS will fill this groove completely.”
Before we go further, there is likely to justify that there is a justification for a brief acting on the “Clade Three Model” names (presented in March 2024). Haiku, Sonnet and OPUS offer a comparison between the price (in API), speed and ability.
Haiku models are smaller and less expensive to operate, and are less capable of what you can call “the depth of context” (taking into account the conceptual relationships in the claim) and encrypted knowledge. Due to its small size in the number of parameters, Haiku models maintain less tangible facts, thus tending to identify more frequently (reasonably answer questions based on data shortage) than larger models, but they are much faster in basic tasks than large models. Sonnet is traditionally a medium model that is a balance between cost and ability, and OPUS models have always been larger and slower. However, OPUS models are more depth and are deeper and are more virtual to operate deep logical tasks.
Claud’s screenshot screenshot with OPUS 4 and SonNet 4 options offered.
Credit: Man
There is no Claude 4 Haiku yet, but the new Sonnet and OPUS models can handle tasks that previous versions have not been able to do so. In our interview with Albert, describe the test scenarios where OPUS 4 worked coherent for up to 24 hours in tasks such as playing Pokemon While coding the expulsion tasks in Claude Code for seven hours without interruption. Albert said, earlier, Claude models continued only an hour before only two hours before losing cohesion, which means that the models can only produce useful self -reference outputs for a long time before starting to make many mistakes.
