GLM 5.2 Dropped Without a Single Benchmark. I Think That's the Whole Point.
On June 13, 2026, Z.ai handed GLM 5.2 to its subscribers. No benchmarks. Not a table, not a single comparison chart. No weights either; the promise was "next week." What did arrive that day was enough to decide whether the model deserved a night of testing: a one-million-token window, up from the previous generation's 200K, which the lab claims holds without degradation, an announced MIT license, and an endpoint that speaks the Anthropic API dialect.
I tend to distrust promotional benchmarks, so the launch won me over before I tested anything. The usual script is predictable: a number leaks, it goes viral, and within 48 hours someone is building a chart against SWE-bench so we can all argue while squinting at graphs. This time the lab skipped that part and handed out access before the numbers. That forces something almost nobody does: judging from the terminal, not from the leaderboard. The price of an informed opinion, this time, was a night on a branch.
I wrote this believing it was Friday. It was Saturday. Doesn't matter, but I'm leaving it in because the whole point of this note is trusting nothing without verification, including my own calendar.
The promise lasted three days
What Z.ai withheld that day was performance. No MMLU, no SWE-Bench, not even a modest internal table. No weights either. "Radical openness" with the repository closed sounded like marketing with an expiration date, and I wrote that at the time: if the repo opens with MIT weights and a serious README, I'll believe them. Until then, it's a promise.
I believed them on Tuesday the 16th, when they opened full access and the weights showed up on HuggingFace (zai-org/GLM-5.2) along with the benchmark table the launch had refused to preview. Seventy-two hours, not a day more. The bet resolved fast, though "radical openness" still strikes me as a grand name for a schedule that was decided long before.
The table brought 62.1 on SWE-Bench Pro, above the 58.4 with which GLM 5.1 had led for months. These are the vendor's own numbers, no independent audit yet. The 58.4 I believe on track record; the 62.1 I'm keeping on a short leash until someone else measures it. What I like about the episode is the lesson more than the number: launching without benchmarks forces you to test with your own work, and that stance lasted exactly as long as a marketing cycle. Three days.
A night on a branch
Before opining I did the only thing that feels honest to me: give it real work.
First task, routine. Moving logic from a controller into a service in a real project, with its own history and conventions. It did well. No invented imports, it respected the repo's style, and the tests came out green. Nothing spectacular, which is exactly what you ask of a second model.
Second task, the one that matters. Chasing a concurrency bug that only appeared under load. That's where it fell short: it got it wrong twice before closing in on the cause. With Opus the search would have been more direct. That's not a cosmetic difference when the bug lives in production; it's the difference between a model you leave alone and one you supervise.
The third was my own context test. The million tokens is its claim and I can't confirm it, because that asks for weeks of use and I had hours. What I did was load a repository, documentation, and issues into a single session until I hit roughly 180k tokens. There it held without losing track of which module was which, something other long-window models have burned me on. The pattern is old: you get past the third file and the model starts mixing what belongs to what. GLM 5.2 didn't trip there, at least not at 180k. From the theoretical million to the usable million there's distance, and the important verb isn't "to have," it's "to use." That part is still pending, mine and everyone's.
MIT and an endpoint I already knew
The switch from Apache-2.0 to MIT looks like a lawyer's detail and it isn't. MIT is the license corporate legal committees sign without arguing; Apache-2.0 drags patent clauses that fuel entire meetings. For a lab that wants its model inside Western companies, simplifying that paperwork weighs as much as any technical feature.
The endpoint is the other half of the play. You point Claude Code or OpenCode at Z.ai's API by changing one environment variable, use the model as GLM-5.2[1m], done. No new SDK to learn. Subscriber quotas, 3× at peak and 2× off-peak, are enough to test with appetite and not enough to move a whole team.
The price got confirmed at 4.40 per million output. Cheap doesn't mean good, but it does change who you hand the routine tasks. The testers on Hacker News place it, on average, 6 to 12 months behind the closed frontier. For daily work, it's enough for me. For the hard stuff, I keep paying the premium.
For now
The verdict hasn't changed since June: GLM 5.2 is a good second model for day-to-day coding, not a replacement for the frontier model when the problem is serious. I haven't migrated anything critical. It stays as my second option in Claude Code for whatever doesn't deserve the expensive model's cost.
And the landscape has already moved again, because nobody rests in this business. On August 14, Z.ai shipped GLM 5.3, with a +50% over 5.2 on its own Z.ai Code Bench. Vendor-reported again, and I haven't touched it. The 5.2 I measured that night is still the same model; what changed is that "the current GLM" is a target that moves every two months. If 5.3 gifts me another free night, it lands on a branch just like its brother. For now, 5.2 stays where I left it: with the tests green.