Headline · l1-l2-bench-v1

DevRev Computer vs Claude Code MCP 2

Same Opus model family. Same tasks. Same judge. Matched on the realistic-API server.

Accuracyi
95.4%
vs 67.9% Claude · +28 pts
Tokens per correct answeri
1.7×
fewer than Claude (428K vs 717K)
Cost as data scalesi
+13%
tokens 1×→256x · Claude +65%
Matched comparison (best config, per scale)
ScaleComputer acc.Claude acc.Δ accuracyTokens / correct
1x96.4%69.3%+27 pts424K / 475K
4x97.1%65.0%+32 pts383K / 667K
16x92.1%70.0%+22 pts419K / 760K
64x96.4%68.6%+28 pts430K / 868K
256x95.0%66.4%+29 pts485K / 817K
Where the gap is

Claude stays close on simpler lookups but falls behind on multi-hop joins and cross-source synthesis (higher task levels).

Token efficiency (tokens per correct)

Lower is better. Computer answers correctly with far fewer tokens at every scale.

Total tokens vs data scale

Computer stays roughly flat as the dataset grows 1×→256x; Claude's token use climbs.

Methodology. Numbers are computed live from pinned canonical runs (latest per agent per scale), not hardcoded. Both sides use the same Opus model family, the same 13-task L1-L2 set, and the same judge. Claude runs use the realistic-API server (MCP2).

Related: ISS-330492 · ISS-330502 · ISS-330503 · ISS-330485.