{"$k":["id","title","columns","rows"],"$r":[["routing-at-scale-inputs","Recorded inputs and cost calculations (per decision)",{"$k":["key","label","unit"],"$r":[["router","Router","text"],["costPer1000","Cost per 1,000 decisions","usd"],["costBasis","Cost basis","text"],["costN","n (cost)","count"],["p50","Median time per decision","ms"],["p95","95th percentile","ms"],["min","Fastest call (range, not an interval)","ms"],["max","Slowest call (range, not an interval)","ms"],["latN","n (time)","count"],["latBasis","Time basis","text"]]},{"$k":["router","costPer1000","costBasis","costN","p50","p95","min","max","latN","latBasis"],"$r":[["Deterministic routing policy",0,"no model call",20000,0.00142,0.00233,null,2.53821,20000,"timed in process (microbenchmark, no database reads)"],["Jev 1.13 (TypeSafe)",0.0337,"provider-reported run cost ÷ calls × 1,000 (calculation)",82,136.5,195.7,100.9,297.3,246,"live run over direct HTTPS, one call at a time (client wall time)"],["Claude Sonnet 5.5 (low) · Claude Code",7.324,"list price × CLI-reported tokens (calculation)",82,2598,4298,1993,5583,82,"recorded routing study, through the Claude Code CLI, one call at a time"],["Claude Haiku 4.5 · Claude Code",8.924,"list price × CLI-reported tokens (calculation)",82,12674,34413,5857,51278,82,"recorded routing study, through the Claude Code CLI, one call at a time (default thinking on)"]]}],["routing-at-scale-formulas","The formulas",{"$k":["key","label","unit"],"$r":[["quantity","Quantity","text"],["formula","Formula","text"],["inputs","Inputs","text"]]},{"$k":["quantity","formula","inputs"],"$r":[["Cost per day","decisions a day × cost per 1,000 decisions ÷ 1,000","Routing receipts: Jev reported total cost ÷ calls × 1,000 (calculation); Claude tokensPerDecision × list prices, with one-hour cache writes"],["Decisions per second","decisions a day ÷ 86,400","A steady rate over 24 hours"],["Decisions in flight","decisions per second × assumed median or p95 time per decision in seconds","Routing-overhead results: policy latencyUs p50 and p95; Claude routing receipts latencyMs wallP50 and wallP95; live Jev run: latencyMs.calls median and p95"],["Waiting hours a day","decisions a day × time per decision in seconds ÷ 3,600","The same times, if each request waits for its decision"],["System One scenario","decisions = model calls a day × (System One decisions per task ÷ model calls per task) = × 7 ÷ 49.5","Routing-overhead results: perTask.systemOneDecisions.median and perTask.modelCalls.median"],["Policy share of one process","decisions per second ÷ measured decisions per second","Routing-overhead results: policy decisionsPerSecond"],["Cost per year","cost per day × 365","A flat daily volume"]]}],["routing-at-scale-context","Counts, inputs and scale checks behind the calculation",{"$k":["key","label","unit"],"$r":[["quantity","Quantity","text"],["value","Value","text"],["n","n","count"],["note","Note","text"]]},{"$k":["quantity","value","n","note"],"$r":[["Cost per 1,000 routing decisions, rule-based policy","$0",20000,"The policy makes no model call."],["Median time per routing decision, rule-based policy","1.42 µs (p95 2.33 µs); maximum 2538.21 µs (minimum unavailable)",20000,"timed in process (microbenchmark, no database reads). The 95th percentile is a point on the time distribution. It is not a confidence interval. The run range, where shown, is not an interval. The minimum and individual timing samples are unavailable; we cannot rebuild the quantiles or full range."],["Median time per routing decision, Jev","136.5 ms (p95 195.7 ms); range 100.9 ms to 297.3 ms",246,"live run over direct HTTPS, one call at a time (client wall time). The 95th percentile is a point on the time distribution. It is not a confidence interval. The run range, where shown, is not an interval."],["Median time per routing decision, Sonnet 5.5 router","2.60 s (p95 4.30 s); range 1.99 s to 5.58 s",82,"recorded routing study, through the Claude Code CLI, one call at a time. The 95th percentile is a point on the time distribution. It is not a confidence interval. The run range, where shown, is not an interval."],["Median time per routing decision, Haiku 4.5 router","12.67 s (p95 34.41 s); range 5.86 s to 51.28 s",82,"recorded routing study, through the Claude Code CLI, one call at a time (default thinking on). The 95th percentile is a point on the time distribution. It is not a confidence interval. The run range, where shown, is not an interval."],["Model calls per task, median of the recorded runs","49.5",48,"Each model call counts as one decision a router could make. Recorded range: 13 to 73 calls; not an interval."],["System One decisions per task, median of the recorded runs","7",48,"Recorded range: 2 to 24 decisions; not an interval."],["System One decisions as a share of model calls, pooled over the runs (calculation)","17.0% (419 of 2,467)",48,"Total System One decisions ÷ total model calls over the runs with at least one call. It differs from the ratio of medians. The scenario chart uses the ratio of medians, as the routing-overhead study does."],["Recorded work cost per model call, median task cost ÷ median calls (calculation)","$0.0611",48,"Notional list-price cost of Agent’s recorded bench work on Sonnet 5.5; the runs used a subscription."],["Decisions in flight at 1 million a day, rule-based policy (calculation)","0.000016 (0.000027 at the 95th percentile time)",20000,"Decisions per second × assumed median or p95 time. These are scenarios, not actual mean concurrency or confidence bounds."],["Decisions per second at 10 million a day (calculation)","115.7 per second",null,"A steady rate over 24 hours."],["Daily cost at 10 million decisions, rule-based policy (calculation)","$0",20000,"$0 a year at a flat volume."],["Daily cost at 10 million decisions, Jev (calculation)","$337.00",82,"$123,005 a year at a flat volume."],["Daily cost at 10 million decisions, Sonnet 5.5 router (calculation)","$73,240",82,"$26,732,600 a year at a flat volume."],["Daily cost at 10 million decisions, Haiku 4.5 router (calculation)","$89,240",82,"$32,572,600 a year at a flat volume."],["Decisions in flight at 10 million a day, rule-based policy (calculation)","0.00016 (0.00027 at the 95th percentile time)",20000,"Decisions per second × assumed median or p95 time. These are scenarios, not actual mean concurrency or confidence bounds."],["Decisions in flight at 10 million a day, Jev (calculation)","15.8 (22.7 at the 95th percentile time)",246,"Decisions per second × assumed median or p95 time. These are scenarios, not actual mean concurrency or confidence bounds."],["Decisions in flight at 10 million a day, Sonnet 5.5 router (calculation)","300.7 (497.5 at the 95th percentile time)",82,"Decisions per second × assumed median or p95 time. These are scenarios, not actual mean concurrency or confidence bounds."],["Decisions in flight at 10 million a day, Haiku 4.5 router (calculation)","1,467 (3,983 at the 95th percentile time)",82,"Decisions per second × assumed median or p95 time. These are scenarios, not actual mean concurrency or confidence bounds."],["Waiting hours a day at 1 million decisions, Haiku 4.5 router (calculation)","3,521 hours (9,559 at the 95th percentile time)",82,"Decisions × assumed median or p95 time. Actual total waiting requires the mean; these are not confidence bounds."],["Work cost of 1 million model calls a day at the recorded cost per call (calculation)","$61,131",48,"A notional list-price cost; your tasks may differ."],["Router cost as a share of the work cost per call, Sonnet 5.5 router (calculation)","12.0%",82,"Cost per decision ÷ recorded work cost per model call, if the router decides every call."],["Router cost as a share of the work cost per call, Haiku 4.5 router (calculation)","14.6%",82,"Cost per decision ÷ recorded work cost per model call, if the router decides every call."],["Router cost as a share of the work cost per call, Jev (calculation)","0.06%",82,"Cost per decision ÷ recorded work cost per model call, if the router decides every call."],["Rule-based policy: measured decisions per second in one batch","469,409 per second",200000,"One process, one 200,000-iteration throughput batch of the routing-overhead study. Database reads and the decision-record write are not included."],["Share of that measured policy rate used at 10 million decisions a day (calculation)","0.025%",null,"Decisions per second at that volume ÷ measured decisions per second."]]}],["routing-at-scale-grid","Every scenario, volume and router (calculation)",{"$k":["key","label","unit"],"$r":[["scenario","Scenario","text"],["calls","Model calls a day","count"],["router","Router","text"],["decisions","Decisions routed a day","count"],["perSecond","Decisions per second (24-hour average)","count"],["costDay","Cost per day","usd"],["costYear","Cost per year (× 365)","usd"],["flightP50","In flight at median time","calls"],["flightP95","In flight at 95th percentile time","calls"],["hoursP50","Waiting hours a day at median time","count"],["hoursP95","Waiting hours a day at 95th percentile time","count"]]},{"$k":["scenario","calls","router","decisions","perSecond","costDay","costYear","flightP50","flightP95","hoursP50","hoursP95"],"$r":[["Every model call routed",10000,"Deterministic routing policy",10000,0.11574,0,0,1.6435e-7,2.6968e-7,0.0000039444,0.0000064722],["Every model call routed",10000,"Jev 1.13 (TypeSafe)",10000,0.11574,0.337,123.01,0.015799,0.02265,0.37917,0.54361],["Every model call routed",10000,"Claude Sonnet 5.5 (low) · Claude Code",10000,0.11574,73.24,26733,0.30069,0.49745,7.2167,11.939],["Every model call routed",10000,"Claude Haiku 4.5 · Claude Code",10000,0.11574,89.24,32573,1.4669,3.983,35.206,95.592],["Every model call routed",100000,"Deterministic routing policy",100000,1.1574,0,0,0.0000016435,0.0000026968,0.000039444,0.000064722],["Every model call routed",100000,"Jev 1.13 (TypeSafe)",100000,1.1574,3.37,1230,0.15799,0.2265,3.7917,5.4361],["Every model call routed",100000,"Claude Sonnet 5.5 (low) · Claude Code",100000,1.1574,732.4,267330,3.0069,4.9745,72.167,119.39],["Every model call routed",100000,"Claude Haiku 4.5 · Claude Code",100000,1.1574,892.4,325730,14.669,39.83,352.06,955.92],["Every model call routed",1000000,"Deterministic routing policy",1000000,11.574,0,0,0.000016435,0.000026968,0.00039444,0.00064722],["Every model call routed",1000000,"Jev 1.13 (TypeSafe)",1000000,11.574,33.7,12301,1.5799,2.265,37.917,54.361],["Every model call routed",1000000,"Claude Sonnet 5.5 (low) · Claude Code",1000000,11.574,7324,2673300,30.069,49.745,721.67,1193.9],["Every model call routed",1000000,"Claude Haiku 4.5 · Claude Code",1000000,11.574,8924,3257300,146.69,398.3,3520.6,9559.2],["Every model call routed",10000000,"Deterministic routing policy",10000000,115.74,0,0,0.00016435,0.00026968,0.0039444,0.0064722],["Every model call routed",10000000,"Jev 1.13 (TypeSafe)",10000000,115.74,337,123010,15.799,22.65,379.17,543.61],["Every model call routed",10000000,"Claude Sonnet 5.5 (low) · Claude Code",10000000,115.74,73240,26733000,300.69,497.45,7216.7,11939],["Every model call routed",10000000,"Claude Haiku 4.5 · Claude Code",10000000,115.74,89240,32573000,1466.9,3983,35206,95592],["Only System One decisions routed (14.1% of calls)",10000,"Deterministic routing policy",1414,0.016366,0,0,2.3239e-8,3.8132e-8,5.5774e-7,9.1517e-7],["Only System One decisions routed (14.1% of calls)",10000,"Jev 1.13 (TypeSafe)",1414,0.016366,0.047652,17.393,0.0022339,0.0032028,0.053614,0.076867],["Only System One decisions routed (14.1% of calls)",10000,"Claude Sonnet 5.5 (low) · Claude Code",1414,0.016366,10.356,3780,0.042518,0.07034,1.0204,1.6882],["Only System One decisions routed (14.1% of calls)",10000,"Claude Haiku 4.5 · Claude Code",1414,0.016366,12.619,4605.8,0.20742,0.56319,4.9781,13.517],["Only System One decisions routed (14.1% of calls)",100000,"Deterministic routing policy",14141,0.16367,0,0,2.3241e-7,3.8135e-7,0.0000055778,0.0000091524],["Only System One decisions routed (14.1% of calls)",100000,"Jev 1.13 (TypeSafe)",14141,0.16367,0.47655,173.94,0.022341,0.03203,0.53618,0.76872],["Only System One decisions routed (14.1% of calls)",100000,"Claude Sonnet 5.5 (low) · Claude Code",14141,0.16367,103.57,37803,0.42521,0.70345,10.205,16.883],["Only System One decisions routed (14.1% of calls)",100000,"Claude Haiku 4.5 · Claude Code",14141,0.16367,126.19,46061,2.0743,5.6323,49.784,135.18],["Only System One decisions routed (14.1% of calls)",1000000,"Deterministic routing policy",141414,1.6367,0,0,0.0000023242,0.0000038136,0.00005578,0.000091526],["Only System One decisions routed (14.1% of calls)",1000000,"Jev 1.13 (TypeSafe)",141414,1.6367,4.7657,1739.5,0.22341,0.32031,5.3619,7.6874],["Only System One decisions routed (14.1% of calls)",1000000,"Claude Sonnet 5.5 (low) · Claude Code",141414,1.6367,1035.7,378040,4.2522,7.0347,102.05,168.83],["Only System One decisions routed (14.1% of calls)",1000000,"Claude Haiku 4.5 · Claude Code",141414,1.6367,1262,460620,20.744,56.325,497.86,1351.8],["Only System One decisions routed (14.1% of calls)",10000000,"Deterministic routing policy",1414141,16.367,0,0,0.000023242,0.000038136,0.0005578,0.00091526],["Only System One decisions routed (14.1% of calls)",10000000,"Jev 1.13 (TypeSafe)",1414141,16.367,47.657,17395,2.2341,3.2031,53.62,76.874],["Only System One decisions routed (14.1% of calls)",10000000,"Claude Sonnet 5.5 (low) · Claude Code",1414141,16.367,10357,3780400,42.522,70.347,1020.5,1688.3],["Only System One decisions routed (14.1% of calls)",10000000,"Claude Haiku 4.5 · Claude Code",1414141,16.367,12620,4606200,207.44,563.25,4978.6,13518]]}]]}