I run a site that tracks which AI model is actually worth paying for, so I check the leaderboards more than most people would want to. Claude Opus 5 sits at #1 out of 215 models on the BenchLM.ai aggregate ranking this week, scoring 85.88 overall, with 96.0% on SWE-bench Verified and 79.2% on the harder SWE-bench Pro. Nothing else on the list is close on raw coding ability. And I still switched most of my own work this week to a model that ranks #4.
That model is GPT-5.6, or more specifically its cheapest tier, which OpenAI calls Luna. Sol, the flagship GPT-5.6 tier, ranks #4 overall at 81.46, with 88.8% on Terminal-Bench 2.1 (91.9% in its ultra mode) and 64.6% on SWE-bench Pro, a real gap behind Opus 5's 79.2%. So on the metric that supposedly matters most for a coding assistant, GPT-5.6 loses outright. I use it anyway for the bulk of my day.
Here's the part the leaderboard doesn't price in: what I actually do most of the time isn't a multi-file refactor or a long agentic tool-use chain. It's small, repetitive, low-stakes stuff. Rename a function across three files. Write a one-off script to reshape a CSV. Draft boilerplate for a new API route I've written thirty times before. For that category of work, the difference between a 96.0% and an 88.8% pass rate on some benchmark essentially never shows up, because the tasks aren't hard enough to separate the two models. What does show up is the bill.
Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, and that's the only tier it comes in. GPT-5.6 splits into three: Sol at $5/$30, Terra at $2.50/$15, and Luna at $1/$6. Luna is roughly a fifth of what Opus 5 costs on the output side. Run enough of those small tasks through an API in a given week and that ratio adds up to real money, not a rounding error.
So my actual routing looks like this. Anything that's genuinely a production PR, a change that touches a lot of files with dependencies between them, or a long tool-use chain where a wrong early step compounds into a wasted afternoon, goes to Opus 5. I've watched Opus 5 hold context through longer agentic chains without losing the plot in a way GPT-5.6 sometimes does past a certain chain length, and for that kind of work the extra cost is trivial next to the time it saves. Everything else, the quick stuff, the stuff where being wrong costs me thirty seconds instead of thirty minutes, goes to Luna or Terra depending on how much I trust the task.
I'll admit the split isn't scientific. I haven't run a controlled study measuring exactly where the quality gap between Opus 5 and Luna starts to matter for my specific workload; I'm going on a few weeks of gut feel and a running mental tally of which model made me redo something. It's entirely possible I'm leaving quality on the table for some subset of "simple" tasks that are actually subtler than I've been treating them. But the leaderboard was never going to tell me that. It ranks models on a fixed set of hard problems, and most of what I actually type into a terminal on a Tuesday afternoon isn't one of them.
If you're trying to figure out your own routing instead of just picking whatever's #1, I wrote up the full breakdown, benchmarks, pricing tiers, and where each model actually wins, at openaitoolshub.org/en/blog/opus-5-vs-gpt-5-6, since both numbers move often enough that a single snapshot goes stale fast.
Top comments (0)