Skip to content
All case studies

CASE STUDY / Coding and repair

Same coding success. 64.4% lower inference cost.

A smaller model matched 95.9% task success across 512 coding and repair tasks, at 64.4% lower inference cost.

64.4%lower inference cost
95.9%task success, matched
512coding and repair tasks

The task: write, test and repair working code.

These tasks require more than a plausible code snippet. The model implements Python, runs checks and repairs failures over a multi-turn episode. Success is determined by executable final checks.

The original GPT-5.4 mini configuration passed 491 of 512 tasks. Isoquant searched for a lower-cost configuration that preserved that level of task success.

A smaller model with a better execution strategy.

Isoquant compared model and instruction combinations, then selected GPT-5.4 nano with a specification-focused checklist. No fine-tuning was needed.

The selected configuration also passed 491 of 512 tasks. Complete-task inference cost, including repair calls, dropped from $4.34 to $1.55 per 1,000 tasks.

The same success rate, with a different cost profile.

The result was 64.4% lower estimated inference cost at the same 95.9% aggregate task success. Mean task completion took 9.31 seconds rather than 6.60 seconds. The objective favored cost while preserving quality, with no latency ceiling.

The task-level results were not identical: 20 previous failures were fixed, and 20 previous successes regressed. Isoquant also retained a higher-quality alternative that reached 98.8% success while still reducing cost by 51.5%.

MATCHED COMPARISON

By the numbers.

BASELINEGPT-5.4 mini
WITH ISOQUANTGPT-5.4 nano + Isoquant optimization
Baseline and optimized results on 512 tasks
MetricBaselineWith Isoquant
Task success95.9%95.9%
Correct / total491 / 512491 / 512
Inference cost / 1,000 tasks$4.34$1.55
Mean completion latency6.60 s9.31 s

More than one way to improve.

These are separate retained configurations, each with its own cost, quality and latency profile.

GPT-5.4 nano + Isoquant optimization, higher-quality alternative

Task success
98.8% (506 / 512)
Cost / 1,000 tasks
$2.11
Mean latency
12.16 s

GPT-4.1 mini + Isoquant optimization, retained alternative

Task success
98.2% (503 / 512)
Cost / 1,000 tasks
$2.47
Mean latency
10.70 s
YOUR WORKLOAD. YOUR SUCCESS CRITERIA.

Let’s build your next performance advantage.

Talk to us
Banking intent classification91.7%lower inference costRead the case study Returns-policy decisions97.5%decision accuracyRead the case study Multi-step order resolution98.7%task successRead the case study