Coding benchmarks show speed and cost separating closely matched models

Leading coding models are delivering increasingly similar results, making speed, pricing, quotas, and tool reliability more important differentiators. In Akita’s llm-coding-benchmark, Gemini 4 Argon detected all 14 planted defects before disclosure and scored 100. The benchmark tests vigilance as agents develop a Rails application over seven sprints, rather than general intelligence or comprehensive programming ability.

GPT-6.1 Terra offered a strong practical balance, scoring 96.8 in 21 minutes at an equivalent cost of $1.69. GPT-6.1 Sol scored 97.0 in 65 minutes. Claude Sonnet 5.5 finished in 27 minutes but left a doctored security test unresolved. Muse Wonder also scored 97.0, although quota interruptions and stalled runs undermined its usability. Its nearly 15-day completion span included quota waits and a voluntary pause.

A provisional Engineering Breadth Index combined the coding benchmark with SWE-bench Pro V2 and Terminal-Bench 4.0. GPT-6 Astra and Gemini 4 Argon effectively tied, both displaying 96.6. Sonnet’s stronger terminal-task performance improved its standing, illustrating how narrow tests can obscure broader strengths. Only 13 models had matching results across the included benchmarks; missing LiveCodeBench results prevented the planned full index.

These rankings remain sensitive to task selection, harness differences, and run-to-run variation. Equivalent costs calculated from token counts also differ from actual subscription bills. Model selection should therefore prioritize performance on real workloads, alongside availability, privacy, operating costs, and the ability to switch providers without losing project history.