Overview

• Anthropic reports the model maintained around 30 hours of uninterrupted, self-directed programming in its internal tests.
• The company says the prior generation sustained similar tasks for about seven hours earlier this year, underscoring a claimed endurance gain.
• Anthropic claims Claude Sonnet 4.5 is particularly strong at spotting exploitable code vulnerabilities and has been used for this in its own development.
• The firm touts leading results on coding and computer-use benchmarks but acknowledges independent evaluations will be needed to verify performance.
• The announcement arrives in a competitive push with OpenAI and Google, with Anthropic indicating a new Opus version is planned later in 2025.