GLM-5.2 crushes open-source coding leaders, lands within 4 points of Claude Opus 4.8

The open-source model race just got a lot more interesting. The team behind GLM has dropped GLM-5.2, and it’s not just another minor point upgrade — it’s a direct shot at the closed-source frontier, specifically targeting long, messy, real-world engineering tasks. The core pitch here is a genuinely usable 1-million-token context window. That’s a claim plenty of models make, but GLM-5.2 backs it up with a new architecture trick called IndexShare that makes the attention mechanism dramatically cheaper to run, slashing per-token FLOPs by 2.9x at that 1M scale. They’ve also tweaked the speculative decoding layer, boosting acceptance length by up to 20%. It’s a package clearly built for the grueling, multi-hour coding agent trajectories where most models start to lose the plot.

On the benchmark scoreboard, the results speak for themselves. On Terminal-Bench 2.1, GLM-5.2 hits 81.0, a massive leap from GLM-5.1’s 63.5 and landing just four points behind the top dog, Claude Opus 4.8. On the ultra-long-horizon SWE-Marathon, it’s still got room to grow — trailing Opus 4.8 by a significant 13% — but it handily beats every other open-source contender and even edges out GPT-5.5 on the FrontierSWE benchmark. The real differentiator might be the new “effort level” control. Instead of a one-size-fits-all thinking budget, users can explicitly dial the model’s compute up or down, letting a developer sprint through simpler tasks with low latency or pour on the gas for gnarly debugging sessions.

Under the hood, those architectural changes are doing the heavy lifting. The IndexShare mechanism reuses a single, lightweight indexer across every four sparse attention layers, which is how they get those computational savings without the model’s attention drifting on long sequences. The MTP layer improvements are similarly clever, fixing a training-inference discrepancy from the previous generation that was quietly leaving performance on the table. It’s the kind of unglamorous systems work that never makes headlines but is the difference between a model that technically supports a 1M context and one that can actually hold a coherent thought across it.

Released under a pure MIT license with no regional restrictions, GLM-5.2 is the highest-ranked open-source model across the board on these long-horizon coding tests. It’s not the absolute king — Opus 4.8 still wears that crown — but the gap is no longer a chasm. For teams that can’t or won’t pipe their proprietary codebase to a closed API, this very suddenly looks like the default option. The question now is whether the next generation of open models can close that final sliver of ground before the closed-source flagships make another leap.