AI BriefingGPT-5.5 Leads in Real Coding Tasks
Key point
Across 56 real coding tasks, GPT-5.5 most consistently passed review.
Details
GPT-5.5, GPT-5.4, and Opus 4.7 were compared on 56 real coding tasks drawn from Zod and graphql-go-tools.
- Each model was run in its own default agent harness.
- The Anthropic line used Claude Code, and the OpenAI line used Codex CLI.
- Evaluation went beyond simple test-pass rates to include parity with the human patch, code review acceptability, patch size/footprint risk, and craft/discipline.
As a result, GPT-5.5 most often came closest to a "deployable" default. It more frequently produced patches that passed tests, better reflected the intended change, and held up under review.
By contrast, Opus 4.7 often wrote smaller patches.
- In Zod, this difference showed up as a real trade-off.
- In graphql-go-tools, smaller patches were seen more often to result in incomplete implementations.
The key takeaway is that no single score can rank all the models. This benchmark shows that the optimal model changes depending on whether the bottleneck is review pass rate or patch footprint.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
New frontier AI models got faster at spatial biology, but accuracy stayed flatGPT-5.6 Sol uses twice the tokens of GPT-5.5