Shift-1 vs. top-model baselines

Live Internal Coding (y-axis zoomed to 80–100%).
Arrow runs from the all-top-model baseline (faded) to Shift-1 (solid).

HARNESSCodexClaude Code

DATASET

ROUTER COST% of baseline

BASELINEperformance

ROUTERperformance

Codex: Sol/Astra router vs. all-Astra

SWE-bench VerifiedPass rate

47%

87.1%

Live Internal CodingUser satisfaction

45%

93.0%

92.4%

Claude Code: Sonnet/Opus router vs. all-Opus

SWE-bench VerifiedPass rate

51%

96.0%

91.1%

Live Internal CodingUser satisfaction

76%

92.2%

90.1%

Router cost shown as % of the all-top-model baseline cost (100%).

Note: there are two reasons Shift-1 is better on Codex than Claude Code: one, Shift-1 is trained/tuned almost entirely on Codex conversations and two, Claude Code has more modest gains on cost due to the extremely short cache timeout (5 minutes) and lower Opus/Sonnet cost differential.