0.4 Points Separate the Top Two Coding Models. Switching Harness Moved Both by Four.

Here is the state of the coding-agent leaderboard as of the beginning of this month. On Terminal-Bench 2.1: GPT-5.6 Sol at 89.5%, Claude Opus 5 at 89.1%, Grok 4.6 at 88.4%. Four tenths of a point between first and second. A publishable result, a dozen LinkedIn posts, and at least one procurement decision.

Now run the same two models on the same benchmark through the Vals AI harness instead: Sol 85.77%, Opus 5 84.64%. Same models. Same benchmark version. Roughly four points lower for both.

The thing separating the two models is 0.4. The thing separating the two harnesses is about 4. The scaffolding is worth ten times the model choice, and the scaffolding is the part nobody puts in the headline.

Read the configuration, not the number

Look at what's in the parentheses on that leaderboard and the illusion of a clean comparison falls apart immediately. Sol's 89.5% is at xhigh effort. Opus 5's 89.1% is with Adaptive Reasoning at Max Effort. Grok's is at high. These are three different amounts of thinking, purchased at three different prices, presented in one column as if they were comparable.

This is the same complaint I had about Gemini 3.5 Flash beating Opus 4.7 at a third of the price, and it hasn't improved in two months. A max-effort score tells you a model's ceiling under ideal conditions. It does not tell you what you'll get at the effort level you can actually afford to run in a loop forty times a day.

GLM-5.3's launch made the same point from the other end this month: matching Opus 4.8's coding score at 50k output tokens instead of 120k. Tokens-per-task belongs on the leaderboard next to accuracy. It almost never is.

The harness is the product

When I built a poor man's agent harness to understand what one actually is, the thing that surprised me was how much of the behaviour lived outside the model. Retry policy. How tool errors get fed back. Whether the loop can see its own previous failure. How aggressively context gets trimmed. Change any one of those and the same model becomes noticeably better or worse at the same task.

A four-point spread between two evaluation harnesses is that effect, measured. Which reframes what the leaderboard is for. It isn't telling you which model to buy. It's telling you that the top three models have converged and the differentiator has moved into the scaffolding — which is the layer you own, and the only layer you can actually improve.

So how should you pick a default?

Not from this. Opus 5 shipped on 24 July at the same price as Opus 4.8 and became the default in Claude Code, which is the decision that actually affects my day — not whether it's 0.4 behind on a table. When I wrote the day-two math on Sonnet 5, the criteria that survived contact with real work were the boring ones, and they still are:

Cost per completed task at the effort level you'll really run, not the one that wins benchmarks. How the model fails — does it stop and ask, or confidently produce something that compiles and is wrong. Tool-call reliability across a long loop, because one bad call at step forty costs more than four points of accuracy. And whether your existing harness, prompts, and permissions transfer without a rewrite.

That last one is why most people won't switch over 0.4 points, and shouldn't. Switching costs a week of retuning to chase a difference smaller than the measurement error between two harnesses.

The verdict

Two things are true at once. Terminal-Bench 2.1 is a genuinely good benchmark and the top-three convergence it shows is real and worth knowing. And the specific ordering at the top is not a decision input for anyone — it's within the noise generated by the harness the evaluation happened to use.

I've been beating this drum since I scorecarded the coding agents, and it rhymes with what I found chasing the 60%-of-code statistic: a single clean number is almost always laundering a pile of methodology you weren't shown. The useful question in August 2026 isn't which model is 0.4 ahead. It's whether your harness is good enough that the difference would even reach your users.

Sources

The 89.5 / 89.1 / 88.4 ordering with effort levels comes from Artificial Analysis's Terminal-Bench v2.1 leaderboard; the lower harness-specific figures are Vals AI's run of the same benchmark. A consolidated August view of both, plus per-task cost, is in Morph's scored leaderboard. If you only click one, make it the Vals page — seeing the same benchmark produce different numbers is the whole argument.