Which AI model is best at coding right now?

Added: 2026-05-21

[00:00:00]I just ran four new models through my coding at Benchmark and here is the updated leaderboard. The coding Benchmark for context test three tasks, a bug fix, a refactor, and a migration task. Each model gets one repair attempt if the first try fails. The primary metric is whether the model posts a clean strict pass with repairs and cost used as a tiebreaker inside each tier.

[00:00:20]The top of the board is a clean full pass cluster. Nine models all posted a three out of three strict pass result.

[00:00:27]GPT 5.5 takes the number one spot and it's the first model in this Benchmark to do it with zero repairs. Every other model in this cluster needed at least one repair. GPT 5.4 drops to number two as a result. Grok 4.3 enters at number three and is a really strong new entrant at just 13 cents. It's the most cost efficient row in the entire clean pass group. And the reason why it comes in at number three behind GPT 5.4 despite being cheaper is that one of the other tiebreakers in this Benchmark is output token efficiency. GPT 5.4 used 6,800 tokens to get the same result that Grok 4.3 did using 13,000 tokens. So that token efficiency is what put GPT 5.4 in second place. If you care more about cost over token efficiency, then Grok 4.3 is definitely competitive and potentially just as good as GPT 5.4 on this Benchmark. We then have another new model, Qwen 3.6 Max Preview at number eight.

[00:01:22]Opus 4.7 then comes in at number nine.

[00:01:25]Technically a clean pass, but it's the only model in the cluster using two repairs. That's why it sits at the bottom of that group. Then we have GLM 5.1 at 10 with a soft pass on format.

[00:01:35]And DeepSeek V4 Pro comes in at number 11. It had two strict passes, but a failed refactor after the repair. So it comes in at the bottom of the coding Benchmark. If you want to see the full breakdown of how this Benchmark works and also where all models rank on the multi-turn Benchmark and the Open Code and Human Eval Benchmarks, click the link below to watch the full video.