A Databricks internal coding benchmark found that AI models with expensive tokens may cost less per task than cheaper models, as price-per-task rather than price-per-token determines true cost. Databricks CTO Matei Zaharia noted that "cheaper per-token does not imply cheaper per-task," citing that Anthropic's Sonnet 5 was around 1.7x cheaper than Opus 4.8 per token but cost $2.09 per task compared to Opus 4.8's $1.94, because Sonnet 5 completed tasks less often (81 percent compared to 87 percent) and used more tokens. Open weight models like Z.ai's GLM 5.2 were "statistically tied with Opus 4.8 on quality, but costing $1.28/task against Opus’s $1.94." Academics had previously observed that in about a third of model comparisons, the cheaper listed model cost more, noting "Gemini 3 Flash's listed price is 80 percent cheaper than GPT-5.4's, yet its actual cost across all tasks is 38 percent higher." The harness software also significantly affected cost-performance; Zaharia stated "the very simple Pi harness got the same success rate as harnesses from the LLM vendors with Opus and GPT 5.5, but at 2x less cost," attributing the difference to context size passed to the model.