Cheap intelligence: routine decisions without long reasoning traces

Routine decisions (expenses, SLA breaches, tickets, ledgers) have short answers. Answer them with reasoning off, let a small decision model score each answer, and escalate only the doubtful ones: 2.5× fewer output tokens and 47% lower cost on 300 back-office tasks.

Hasan Shabanideepseek/deepseek-v4-flash-0731300 back-office tasks660 benchmark questions

2.5×
fewer output tokens than always reasoning
314 vs 793 per task
47%
lower cost
$0.070 vs $0.132 per 1,000 tasks
96.3%
accuracy
against 99.7% with full reasoning

Method

Questionone promptFast answerreasoning offDecision modelP(correct)≥ 0.7Submitdone< 0.7cheap retry: re-scoreReasoning passor another cheap trypaid only when the score is low
Figure 1. Reason only below 0.7. A cheap retry is scored again (dashed).

Below 0.7, gate + reasoning runs the reasoning pass; gate only retries cheaply and keeps the best-scored answer.

Back-office decisions

WorkloadGateReasoningOutput tokensSavingEscalated
Reconciliation100%100%12 vs 1159.6×4%
Expense rulings100%100%10 vs 969.3×8%
SLA breach calls98%100%48 vs 1763.7×20%
Ticket routing100%100%22 vs 713.2×18%
Contract dates84%98%1,519 vs 3,7392.5×20%
Invoice arithmetic96%100%270 vs 5582.1×0%
All 300 tasks96.3%99.7%314 vs 7932.5×12%

96.3% against 99.7% (McNemar p = 0.009). Shaded rows: both score 100%. Reasoning still pays off on contract dates (98% against 84%).

General benchmarks

No reasoningGate onlyGate + reasoningFull reasoning

73% 78% 83% 89% 94% 0 1,000 2,000 3,000 4,000 5,000 tokens per question → ← accuracy dashed: split the traffic between the two baselines No reasoning77.9% · 822 tokNo reasoning: 77.9% accuracy, 822 tokens per question, n=660 Gate only84.2% · 1,433 tokGate only: 84.2% accuracy, 1,433 tokens per question, n=660 Gate + reasoning87.6% · 3,522 tokGate + reasoning: 87.6% accuracy, 3,522 tokens per question, n=660 Full reasoning88.5% · 4,465 tokFull reasoning: 88.5% accuracy, 4,465 tokens per question, n=660
Figure 2. Accuracy against tokens on 660 held-out questions. Gate only sits +4.6 points above the dashed traffic-split baseline (95% CI +1.8 to +7.5).
PolicyAccuracy95% CITokensvs reasoningTokens / correct
No reasoning77.9%75%–81%822-10.6 pts · 0.18×1,056
Gate only84.2%81%–87%1,433-4.2 pts · 0.32×1,701
Gate + reasoning87.6%85%–90%3,522-0.9 pts · 0.79×4,021
Full reasoning88.5%86%–91%4,465reference5,047

No reasoningGate onlyGate + reasoningFull reasoning

GPQA 68% · 838 tokNo reasoning on GPQA: 68.2%, 838 tokens, n=198 78% · 2,067 tokGate only on GPQA: 78.3%, 2,067 tokens, n=198 86% · 6,674 tokGate + reasoning on GPQA: 85.9%, 6,674 tokens, n=198 84% · 7,658 tokFull reasoning on GPQA: 83.8%, 7,658 tokens, n=198 MMLU-Pro 78% · 646 tokNo reasoning on MMLU-Pro: 78.0%, 646 tokens, n=200 84% · 1,081 tokGate only on MMLU-Pro: 84.5%, 1,081 tokens, n=200 84% · 2,781 tokGate + reasoning on MMLU-Pro: 84.5%, 2,781 tokens, n=200 84% · 3,469 tokFull reasoning on MMLU-Pro: 84.0%, 3,469 tokens, n=200 MATH-500 91% · 941 tokNo reasoning on MATH-500: 91.0%, 941 tokens, n=134 90% · 1,168 tokGate only on MATH-500: 89.6%, 1,168 tokens, n=134 93% · 1,113 tokGate + reasoning on MATH-500: 92.5%, 1,113 tokens, n=134 95% · 2,536 tokFull reasoning on MATH-500: 94.8%, 2,536 tokens, n=134 AIME 2025 43% · 2,783 tokNo reasoning on AIME 2025: 43.3%, 2,783 tokens, n=30 70% · 3,556 tokGate only on AIME 2025: 70.0%, 3,556 tokens, n=30 77% · 8,370 tokGate + reasoning on AIME 2025: 76.7%, 8,370 tokens, n=30 93% · 10,753 tokFull reasoning on AIME 2025: 93.3%, 10,753 tokens, n=30 Stress suite 90% · 389 tokNo reasoning on Stress suite: 89.8%, 389 tokens, n=98 93% · 581 tokGate only on Stress suite: 92.9%, 581 tokens, n=98 94% · 474 tokGate + reasoning on Stress suite: 93.9%, 474 tokens, n=98 97% · 761 tokFull reasoning on Stress suite: 96.9%, 761 tokens, n=98
Figure 3. By benchmark. The gate gives up most on AIME 2025 (n = 30), competition maths that needs a long trace.

Limitation

The model was never trained to work with a gate: rejected answers mostly come back unchanged, so gating every step of a long task falls apart.

No reasoning73% solved · 29,042 tokNo reasoning: 73.3% of 15 episodes solved, 29,042 tokens each Gate only53% solved · 105,832 tokGate only: 53.3% of 15 episodes solved, 105,832 tokens each Gate per step (veto)87% solved · 33,076 tokGate per step (veto): 86.7% of 15 episodes solved, 33,076 tokens each Full reasoning100% solved · 64,212 tokFull reasoning: 100.0% of 15 episodes solved, 64,212 tokens each
Figure 4. Gating every step of 15 long-horizon episodes: gate only 53%, below no reasoning (73%). Vetoing only confident rejections recovers to 87%.

Next step

Post-train the model on this run's graded records so it defers and revises by itself, then re-run the same tasks.