GLM-5.2 closes the gap with the best closed model to 0.7 points on FrontierSWE: an MIT-licensed coding model

In coding AI, an open-weight model has pulled up right behind the best closed model. GLM-5.2, released by China's Zhipu (Z.ai) on June 13, 2026, scored 75.1% on FrontierSWE, just 0.7 points behind Claude Opus 4.8 at 74.4%, and its weights ship under the MIT license. Its output price is about one-sixth of Opus 4.8. The top open-source coding model is thinning the moat around closed models. ASAP summarizes the result from the primary announcement.

What was actually released

Zhipu opened GLM-5.2 in full on June 13, 2026 and published the weights under the MIT license. Anyone can download the model, self-host it, and use it commercially without restriction. Top-tier coding performance that used to live only behind a closed API has moved into open weights. On specs, it supports a 1 million token context and up to 128K tokens of output, and it offers two reasoning modes with adjustable thinking effort. Zhipu defines it as a base model for agentic engineering, built to handle long codebases and multi-step tasks in one pass.

The ranking across three benchmarks

Performance splits across three measures. On FrontierSWE it scored 75.1% to the best closed model, Claude Opus 4.8, at 74.4%, a 0.7-point gap; on MCP Atlas it was 77.8% to 77.0%, a 0.8-point gap. On Code Arena Frontend, a blind evaluation with a million users, it reached an Elo of 1,595, first among available models and second overall behind the now-suspended Fable 5. Across all three, GLM-5.2 ranks first among open-source models for coding.

How to read that 0.7-point gap

On its own, 0.7 points looks small enough to be noise. But the weight of that gap is set by price and license. The point is that an open-weight model, cleared even for commercial use, lands within a single point of the best closed one. The real decision question shifts from "which is number one" to "can I accept a gap this size at one-sixth the price." For most coding workloads, 0.7 points is a margin humans can close with review and testing. On the other hand, in autonomous agent loops that run without a human in the loop, a small gap can compound into a higher cumulative failure rate, so the weight of it depends on the use case.

Where the unit economics change

GLM-5.2 is priced at roughly $4.40 per million output tokens, about one-sixth of Claude Opus 4.8's $25. Input is also low at about $1.4 per million tokens. This structure matters because coding is an output-heavy task. As long refactors, bulk code generation, and repeated attempts pile up, the output rate drives total cost. When performance is within 0.7 points but output costs one-sixth as much, the habit of defaulting to the most expensive model becomes a cost in itself. Picking a model per task, and attaching a cheaper one to verifiable coding, is what separates the bill.

What Korean practitioners should weigh

Korean teams face a two-way call. First, because it is MIT-licensed, self-hosting it on in-house GPUs lets you run top-tier coding performance without sending data outside. The more sensitive the internal code, the safer this path is over an API. Second, if you use it via API for convenience, you have to weigh questions about China-based data handling. In other words, GLM-5.2's value is not just "it's cheap" but that organizations under tight regulatory and security requirements can reclaim control through open weights.

The question marks that remain

GLM-5.2's benchmark numbers are still Zhipu's own claims and have not been independently verified by a neutral harness. Self-reported scores may have been measured under favorable conditions, so it is reasonable to wait for reproduction on third-party leaderboards. The 0.7-point conclusion itself hangs on that verification. Rather than rushing adoption, running your own A/B on representative tasks to measure the gap on real workloads is the safer move.

Where this release points

GLM-5.2 shows that the lead of closed models in coding is narrowing fast. When an open-weight model lands within a single point of the best at one-sixth the price, and under a commercially free MIT license at that, the moat around closed models shifts away from performance and toward trust and verification infrastructure. The next axis of competition is likely to move from "who is smarter" to "who is cheaper and more controllable."

Source: ASAP summary of Zhipu (Z.ai)'s GLM-5.2 release (June 13, 2026; MIT license, 1M context and 128K output, 75.1% on FrontierSWE vs Claude Opus 4.8's 74.4% for a 0.7-point gap, 77.8% vs 77.0% on MCP Atlas, first among available models on Code Arena, about $4.40 per million output tokens vs Opus 4.8's $25) and related reporting.