Recently, within a month, we've been amazed by all these new AI models from three giants: OpenAI, Google, and Anthropic!
From Anthropic, we have Claude Opus 4.5 (with the highest SWE Benchmark 80.9%); from OpenAI, their flagship coding model GPT-5.2 (Codex) with SWE Benchmark 80%; and from Google, Gemini 3 Pro, which at launch was said to be SOTA in most benchmarks and boasts advanced agentic capabilities.
The catch here is that there are so many models available for coding or agentic coding that it's hard to decide which one to pick as your daily driver.
All of them claim at some point to be the "so-called" best for coding. But now the question arises: in actual agentic coding, which involves working on a production-ready project, how much better or worse is each of them in comparison?
TL;DR
If you want a quick take, here’s how the three models performed in these two tests:
Claude Opus 4.5: Safest overall pick from this run. It got closest in both tests and shipped working demos, even if there were rough edges (hardcoded values, weird similarity matching). Also, the most expensive.
Gemini 3 Pro: Best result on Test 1. The fallback and cache were actually working and fast. Test 2 was weird; it kept hitting a loop, which resulted in halting the request.
GPT-5.2 Codex: Turned out to be the least reliable for me in these two tasks. Too many API and version mismatches, and it never really landed a clean working implementation.
One thing I really hate about Opus in Claude Code is how much it does web searches. Across these two tests, it did like 30+ web searches, which ended up eating up a big part of the total time. Web search is great, but it gets super frustrating fast when you have to keep approving it and typing “Yes” over and over.
Has anybody else felt this with Claude Code, especially Opus 4.5?
⚠️ NOTE: Don’t treat these as a hard rule. This is just two real dev tasks in one repo, and it shows how each model did for me in that exact setup.