How to compare coding models with Unknown

Kilo Code's authors compared GPT-5.2 and GPT-5.2 Pro with GPT-5.1, Claude Opus 4.5 and Gemini 3 using three coding tasks from an earlier comparison. Each test started from an empty project in Kilo Code. They used Code Mode for the first two tests and Ask Mode followed by Code Mode for the notification-system extension. For the two GPT-5.2 models, they selected the highest available reasoning effort.

The first task asked for a Python TokenBucketLimiter with ten requirements, including exact signatures, an error-message format, time.monotonic() and threading.Lock(). The authors reported that GPT-5.2 Pro alone achieved a perfect score. Its request took seven minutes, and it matched the specified constructor and internal token naming. They also reported that GPT-5.2 avoided the unrequested exceptions GPT-5.1 introduced.

The second task supplied a 365-line TypeScript API handler with SQL injection vulnerabilities, hardcoded secrets and missing validation. The requested refactor separated Repository, Service and Controller layers, added Zod validation and included rate limiting. Both GPT-5.2 variants implemented rate limiting; the source says GPT-5.1 missed it. Opus 4.5 added rate-limit headers and expired-entry cleanup. The supplied excerpt describes the third task's setup but gives no extension result.

Who it's for: Developers comparing coding assistants against explicit implementation requirements.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: dev.to · kilocode