One of the authors here - We had both run through a curated set of 64 long horizon coding tasks to evaluate cost/intelligence. Astra surprised us in this default setting. How is the rest of you faring (especially interested in those who use both)
Context : We continuously do runs of a subset of 64 curated SWE-Bench Pro long horizon tasks on various models. Some via their respective API's. Some on reference hardware setups. We do this to approximate 'real world use' of these models and get more insights on how these use the underlying compute (we advise HW builders)
In our runs we create multiple separate sandboxed agents that use a given model/harness combo (in this case just the default harnesses for both) and feed them the benchmark tasks, which we compare against the golden resolution. (There's also a time out, just to make sure we don't blow through our entire budget by accident). You pay for the tokens no matter if the task gets solved or not, so both bills cover all 64 attempts, also the failed ones (30 for Astra, 13 for Fable). The €68 is just the difference between the two full runs, not the price of those 17 tasks. That's basically why we look at cost per solved task, €1.6 vs €2.4. But we see how the wording of that sentence wasn't optimal. We'll fix that bullet (thx)
Aside from it giving us a good view on evolving token needs of different model generations (and being able to compare with open weights models), the variations in "intelligence per dollar" per generation is also pretty interesting. Hence posts like these, just to share the data, which is hopefully useful to others too. And : always interested in seeing different results.
One of the authors here - We had both run through a curated set of 64 long horizon coding tasks to evaluate cost/intelligence. Astra surprised us in this default setting. How is the rest of you faring (especially interested in those who use both)
So Astra billed for tasks it couldn’t solve?
Correct. But that's by design here.
Context : We continuously do runs of a subset of 64 curated SWE-Bench Pro long horizon tasks on various models. Some via their respective API's. Some on reference hardware setups. We do this to approximate 'real world use' of these models and get more insights on how these use the underlying compute (we advise HW builders)
In our runs we create multiple separate sandboxed agents that use a given model/harness combo (in this case just the default harnesses for both) and feed them the benchmark tasks, which we compare against the golden resolution. (There's also a time out, just to make sure we don't blow through our entire budget by accident). You pay for the tokens no matter if the task gets solved or not, so both bills cover all 64 attempts, also the failed ones (30 for Astra, 13 for Fable). The €68 is just the difference between the two full runs, not the price of those 17 tasks. That's basically why we look at cost per solved task, €1.6 vs €2.4. But we see how the wording of that sentence wasn't optimal. We'll fix that bullet (thx)
Aside from it giving us a good view on evolving token needs of different model generations (and being able to compare with open weights models), the variations in "intelligence per dollar" per generation is also pretty interesting. Hence posts like these, just to share the data, which is hopefully useful to others too. And : always interested in seeing different results.
I’d rather pay more and get things solved.
Penny wise and pound foolish comes to mind.