Databricks reports 60% higher coding spend on the model benchmarks score as cheaper

Cost per task is a unit price. The invoice is that price multiplied by the number of tasks, and engineers set the second number.

The two Databricks figures only sit together if the work got bigger. According to the AI News roundup, @pwendell reported Astra outperforming prior top-end models on complex, long-horizon tasks while coding spend rose about 60 percent, across a rollout to roughly 3,500 engineers ^[4]. The plausible path is that engineers started handing it the jobs the previous model could not finish.

The roundup does not include a baseline dollar figure, a time window, or the size of Astra's per-task discount ^[7], so the volume implication has to be written as a function of that discount: spend at 1.6 times the old level with a per-task price 10 percent lower implies 1.78 times as many task-equivalents, and a 25 percent lower per-task price implies 2.13 times ^[8]. On either reading the buyer is consuming close to twice the agent work and paying 60 percent more in total.

The number rides on one post, summarized in a newsletter, and the newsletter describes it two ways: as plus 60 percent overall spend ^[1] and as coding spend up about 60 percent ^[4]. Those are different denominators. AI News also hedged the benchmark comparison directly, writing that Astra "is not universally cheaper everywhere" ^[3] while noting that many benchmarks score it cheaper than Sol on cost per task because of token efficiency ^[2].

The increase could be a rollout artifact, with 3,500 engineers testing a new top-end model at once. Or price competition erases the problem: Cline has made Union Alpha free, claiming coding performance near GPT-6 Astra and Opus 5 at far lower cost ^[6], and if that claim survives contact with production work then per-token prices fall faster than usage climbs. Or the extra spend bought proportionally more finished work, which nobody in this issue measured. The only output figure in the same roundup is Steve Yegge's, who shut down Gas Town and acknowledged that after spending many thousands a month on coding agent subscriptions, Gas Town was the only thing he ever built with it ^[5].

In my view the narrow conclusion holds: a cost-per-task table tells a buyer which model to test. It does not tell them what the bill will be, because the quantity moves when the price moves. Two disclosures would break that. If Databricks says the 60 percent includes more seats or a longer measurement window, the increase is accounting. If coding spend drifts back toward the old baseline once the new model stops being new, then this was a spike in curiosity.