Cortex AISQL’s 8× cost result is a model of a system—not a measured system result

A new Cortex AISQL paper offers a useful cost model—and an easy number to overread. Its headline result is a roughly 8× total-cost reduction from combining online filter ordering with proxy-to-oracle cascade routing. But the paper’s own evaluation makes the status precise: 7.8× is an analytical estimate for a representative workload, not an end-to-end measurement of the two components running together at production scale. (paper, Sections 6.3–6.4)

What the calculation assumes

The modeled workload contains 1 million rows and a conjunction of five semantic predicates. The calculation assumes filter ordering cuts expected predicate evaluations from five to 1.25 per row, while the cascade sends 30% of evaluated predicates to an oracle model and charges the proxy at 5% of oracle cost. Under independence, those factors reduce modeled cost to 8.75% of baseline, an 11.4× upper bound. (paper, Section 6.3)

Those inputs are anchored in separate component results. The Larch-Sel ordering method reports typical 4×–8× token-overhead reductions on a mixed AND/OR workload, with a 19× ceiling against Palimpzest and Quest. GAMCAL reaches F1 ≥ 0.95 on six classification, filtering and join benchmarks, requiring 17%–68% oracle delegation; it uses up to 58% fewer oracle calls than LOTUS’s single-pass cascade and leads on five of six benchmarks. Those are measured component-level findings, not a measured composed run. (paper, Sections 4–5)

Why the estimate falls to 7.8×

The paper then models two steady-state interactions. Rows surviving earlier predicates can be harder for the cascade, increasing delegation; proxy decisions can also inject noisier labels back into the filter-ordering learner, weakening short-circuit selection. At midpoint penalties of 1.25× and 1.2×, respectively, the estimate falls from 11.4× to about 7.8×. Its pessimistic corner is 5.7×; the optimistic corner returns to 11.4×. A third cold-start penalty is omitted from the long-run figure because the paper estimates it affects roughly the first 1% of batches for the million-row setup. (paper, Section 6.4)

The distinction matters because “production case study” describes Cortex AISQL as the substrate, while the combined number remains a model assembled from earlier component evidence. The paper explicitly names empirical validation of the composition argument at scale as future work. (paper, Sections 3 and 9)

For practitioners, 7.8× is therefore a planning hypothesis with declared parameters—not a benchmark result to copy into a capacity forecast. A defensible reproduction should report the actual row count, predicate count, proxy/oracle cost ratio, delegation rate, quality target, and end-to-end cost with both learners enabled. Until that experiment exists, the paper’s most valuable contribution is not the multiplier; it is the accounting that shows exactly where the multiplier can shrink.