Matthew Berman compares Anthropic's reported Opus 5 results with Fable 5 and earlier Opus models. Coding, computer useComputer use lets an AI agent interact with software through interface controls such as screens, buttons, keyboards and pointers. and interactive-game scoresA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. improve in the presented charts, while some legal and health measures decline.
His central point is to compare successful task costCost per completed task measures the total AI, tool, infrastructure, retry, and repair expense for each verified useful outcome., not just token pricing. Reasoning settingsReasoning effort is the amount of internal computational work an AI model applies before producing an answer or action. can change both cost and accuracy, and the most expensive setting does not always produce the strongest result. He highlights a reported ARC-AGI 3 jump but does not present a comprehensive independent test.
Matthew Berman also discusses launch pricing and automatic safety-triggered model fallbacks. His explanation of reduced cybersecurity capability and possible government review is explicitly speculative, rather than a verified account of training or approval.
Watch on YouTube




