An Open-Weight Model Briefly Tops the Hardest Coding Benchmark

The leaderboard at the top of agentic coding benchmarks has been a closed-weight club for as long as the benchmarks have existed. That changed for nine days in April.

Z.ai's GLM-5.1, released as open-weights on 8 April 2026, posted 58.4 percent on SWE-bench Pro, becoming the first openly licensed model to top the leaderboard. The lead held until Anthropic's Claude Opus 4.7 arrived on 16 April and pushed the bar to 64.3 percent. Brief as it was, the result mattered, because the model is genuinely available to download, fine-tune, and run.

Why the benchmark matters

SWE-bench Pro tests whether a model can resolve real GitHub issues end-to-end, from reading the codebase to writing a patch that passes the project's own tests. It is harder than the older SWE-bench because the tasks were curated to defeat the kind of shortcut-learning that boosts scores on simpler tests. A model that does well on it can do useful engineering work without a human babysitting each step.

Until this spring, every public score above 50 percent on the test had come from a closed system. GLM-5.1 putting an open model above that line, however briefly, is the kind of signal that changes procurement decisions inside large engineering teams.

The shape of the model

According to a detailed comparison in late April, GLM-5.1 is a mixture-of-experts model that favours strong tool-use and structured-output behaviour over raw chat fluency. Z.ai positions it as a coding and agent model rather than a general assistant, and the SWE-bench Pro result is consistent with that focus. A separate broader evaluation noted that GLM-5.1 drops to mid-tier on certain coding methodologies, so the headline number does not generalise to every kind of engineering task.

The release sits inside a wider cluster. DeepSeek V4 and Moonshot's Kimi K2.6 both landed in the same fortnight, alongside MiniMax M2.7. The cumulative effect is that four open-weight Chinese systems are now plausibly substitutable for closed Western models on most coding workloads, at meaningfully lower cost per token.

What it does not change

The closed frontier still leads on harder reasoning, longer planning horizons, and the kind of reliability benchmarks that have lagged the headline scores. Claude Opus 4.7's lead on SWE-bench Pro reasserted itself within a week, and the gap on more demanding tasks is wider than the leaderboard alone suggests. Open-weight models also carry the usual operational tax: someone has to host them, monitor them, and patch them.

What changes is the floor. A team that needs strong coding assistance no longer has to choose between paying frontier prices or accepting a real capability gap. There is now an in-between option that you can run on your own hardware, at your own latency, on your own data, with the weights in your own bucket. That is the real news, even if Anthropic took the top spot back nine days later.