Newer Claude models use more tokens but cost less per task solved (signoz.io)

Recent evaluations of Claude's latest models, particularly Claude Opus 4.8, indicate significant improvements in efficiency and task success while reducing costs per task. The analysis, using a public benchmark called Terminal-Bench, demonstrated that while Opus 4.8 handled an increased number of tokens compared to its predecessors, it ultimately delivered better performance at a lower cost per completed task. Specifically, Claude Opus 4.8 solved eight out of ten tasks, with a cost of approximately $1.01 per task, outperforming Sonnet 4.6 and Opus 4.7. This development is noteworthy for the AI/ML community as it challenges traditional assumptions that newer models will always be more expensive or less efficient. The findings highlight the necessity of detailed performance metrics, such as task completion rates and token utilization breakdowns, rather than relying solely on aggregate scores or token counts. Claude's approach, which emphasizes operational efficiency through advanced telemetry, suggests that future model evaluations should incorporate nuanced performance measures to truly assess their value, shaping how developers and businesses choose AI solutions for complex problem-solving tasks.