Example trace
Every request in a real conversation, priced twice. Once as billed, and once as the same request would have cost on Astra with the same tokens. Click any request on the timeline to see what happened there.
SignsSigns show the effect on cost. Savings are negative and losses positive, and starting from the all-Astra cost they add up exactly to what was billed.
The solid line is what was billed, coloured by the model that served each request, with Astra stretches shaded. The gap to the dashed line is the money saved so far, and the line jumps at each cold write. Click the chart to select a request.
Billed cost minus all-Astra cost for each request. Bars below zero are savings, which is almost every Sol request. Bars above are losses from cold writes after a switch. Astra requests on a warm cache cost exactly the baseline and sit at zero.
Every request is priced twice, once as billed and once as the same request would have cost on Astra with the same tokens, which is the all-Astra baseline. Savings are what Sol’s lower cache read, cache write and output prices take off while caches are warm.
Losses are what a switch costs. The model that takes over has to write context it would otherwise have read from its cache, which means an Astra re-warm when the router escalates and a Sol cold start when it hands back. A gap long enough to expire every cache costs the all-Astra run the same cold write, so it is not counted against the router.