GLM-5.2 Benchmarks: Scores, Sources and Caveats

GLM-5.2, Z.ai's June 2026 flagship, scores 62.1 on SWE-bench Pro, 81.0 on Terminal Bench 2.1 with Terminus-2 (82.7 with the best harness), 46.2 on DeepSWE and 91.2 on GPQA-Diamond, all reported by Z.ai on its own model card. The weights are MIT-licensed with a 1M context. GLM-5.3 has since replaced it as the flagship. Checked 31 October 2026.

Where these numbers come from

Every figure on this page is vendor-reported. The main source is the GLM-5.2 model card on Hugging Face, published 16 June 2026. Where we quote a later number we name the card: the GLM-5.3 card, DeepSeek's V4 Pro card or Moonshot's Kimi K3 card, each of which lists GLM-5.2 as a competitor. We have not reproduced any score, and we did not find an independent leaderboard covering all of these tests.

The card's closed-model columns carry the same caveat. Z.ai marks results with an asterisk where they come from the full set of a benchmark rather than a text-only subset, and it chose which settings to run for rivals.

The full table from Z.ai's GLM-5.2 card

Reasoning

Coding

Agentic

How to read the table

The jump from 5.1 to 5.2 is large. DeepSWE moved from 18 to 46.2, Terminal Bench 2.1 with Terminus-2 from 63.5 to 81.0, and FrontierSWE from 30.5 to 74.4. Those are the rows where Z.ai's long-horizon training should show, and they do.

GLM-5.2 led the open models on those rows at the time. On the card's own columns, it beats DeepSeek-V4-Pro and Qwen3.7-Max on SWE-bench Pro, NL2Repo, DeepSWE and Terminal Bench. It trails Claude Opus 4.8 on SWE-bench Pro (62.1 against 69.2), NL2Repo (48.9 against 69.7) and DeepSWE (46.2 against 58). On math it is close to or ahead of the closed models: AIME 2026 at 99.2 beats the 98.3 listed for GPT-5.5.

It is behind on tool-heavy tasks. On Tool-Decathlon GLM-5.2 is 48.2 against 52.8 for DeepSeek-V4-Pro and 59.9 for Opus 4.8.

Open-weights-only rivals are mixed. On GPQA-Diamond, MiniMax M3 is listed at 93 against GLM-5.2's 91.2, and on HMMT Feb 2026 Qwen3.7-Max is at 97.1 against 92.5.

How the tests were run

The card's footnotes are the part worth reading, because settings change scores. A summary of the stated conditions:

Two things stand out. First, the 400K context windows on several coding tests are well under the 1M the model supports, so the card's coding numbers do not test the full window. Second, three of the rows (FrontierSWE, PostTrainBench and SWE-Marathon) were run by outside groups, which is closer to an independent result than the rest, although Z.ai still reports them.

How the numbers moved later

Scores for the same model change between cards because benchmarks update and harnesses differ. Comparing GLM-5.2's figures across three later cards shows how soft they are.

DeepSeek's V4-Pro card also lists GLM-5.2, giving 81.0 on Terminal Bench 2.1, 46.2 on DeepSWE and 54.7 on HLE with tools, which match. Its NL2Repo figure for GLM-5.2 is 48.9, matching again. The rows that moved are the ones with live leaderboards or revised metrics. For decisions, use the stable rows and do not mix rows from different dates.

What GLM-5.2 is, besides a score

The card describes GLM-5.2 as "our latest flagship model for long-horizon tasks", the first GLM with what Z.ai calls a solid 1M-token context. Other documented features:

• Flexible thinking effort. Multiple levels to trade latency for quality.
• IndexShare. A technique that reuses the same indexer across every four sparse-attention layers, which Z.ai says cuts per-token compute 2.9 times at 1M context.
• Better multi-token prediction. Z.ai says the speculative decoding acceptance length improved by up to 20 percent.
• MIT licence. "No regional limits", in Z.ai's phrase.

It is a text-only model, as is GLM-5.3, which shares its base. Its weights are about 753 billion parameters, so it needs a multi-GPU server. The serving versions listed on the card were SGLang 0.5.13.post1 or later, vLLM 0.23.0 or later, Transformers, KTransformers and Unsloth, plus Ascend NPU frameworks.

What Z.ai's developer docs add

Z.ai's GLM-5.2 documentation page repeats the card's headline numbers and adds claims in prose. It says GLM-5.2 is the strongest open-source model on standard coding benchmarks and, on Terminal-Bench 2.1, lands "within a few points of Claude Opus 4.8" at 81.0 against 85.0. It says that on FrontierSWE, PostTrainBench and SWE-Marathon the model is the highest-ranked open-source one, trailing Opus 4.8 by about one percent on FrontierSWE, and ahead of GPT-5.5 and Opus 4.7 on several tests.

One detail does not line up. The docs page gives GLM-5.1's Terminal-Bench 2.1 score as 62.0, while the model card's table shows 63.5. A 1.5-point difference between two pages from the same maker is small, and probably reflects an update to one of them, but it is a reminder to cite the page and date with any score. The docs also pitch practical examples: a level-based mini game, reproducing a research paper as runnable PyTorch, and a code-to-video loop using Remotion. Those are demonstrations, not measurements.

The gap to closed models, in points

Subtracting within the GLM-5.2 card gives a plain picture of where the model stood in June against the closed models Z.ai chose to list. These are our subtractions of vendor figures.

The pattern is consistent. GLM-5.2 was close or ahead on math and on long-horizon and terminal tasks, and well behind on repository-scale tests such as NL2Repo and DeepSWE. If your work is single-repository feature building, the repository rows matter most to you. If your work is long autonomous terminal sessions, the terminal and FrontierSWE rows matter more.

A test plan that does not need benchmarks

If you are deciding whether GLM-5.2 or its successor suits your work, a day of testing says more than any table.

1. Choose 15 to 20 real tasks from your own history. Include some that failed before: flaky tests, an awkward refactor, a bug with a long trace.
2. Run each task through the same agent with GLM-5.3 and one other candidate, at the same effort level. GLM models default to max effort unless you set it, so state the level on both sides.
3. Record pass or fail, wall-clock time, output tokens and cost per task. Cost per passed task is the figure to compare.
4. Repeat each task at least twice. Agent runs vary.
5. Keep a closed model as a baseline if your policy allows it.

You can run this through Z.ai's API or the Coding Plan with an agent such as Claude Code, OpenCode or Cline. The cost side is simple arithmetic from the prices above: at $1.40 and $4.40 per million tokens, a run of 3 million input and 300,000 output tokens costs $5.52 before cache discounts.

Prices and access today

GLM-5.2 is still on Z.ai's API at $1.40 input and $4.40 output per million tokens, with cached input at $0.26, the same as GLM-5.3. On the Coding Plan, requests for GLM-5.2 are routed to GLM-5.3. The weights stay on Hugging Face under MIT, so self-hosting a frozen GLM-5.2 remains possible for anyone who needs a fixed model.

See GLM models for the current lineup, the Coding Plan and the GLM-5.3 scores, and best open source LLM for coding for the cross-lab view.

Should anyone still pick GLM-5.2?

Checklist for reading any GLM benchmark claim

1. Find the harness. Claude Code, OpenHands, Terminus-2 and mini-swe-agent give different numbers on the same model.
2. Find the effort level. GLM models default to max when unspecified.
3. Find the context length used. Several coding tests ran at 400K, not 1M.
4. Check who ran it. Maker-run, outside-run and leaderboard results are different grades of evidence.
5. Check the benchmark version and date. Terminal Bench 2.1, 3.0 and 4.0 are different tests.
6. Rerun a sample on your own repository before you trust it.

Our take

GLM-5.2's card is a good example of a well-documented vendor table: full footnotes, stated harnesses and a few results from outside groups. It also shows why we call these figures a shortlist. The same model scores 63.7 or 9.5 on ProgramBench depending on which metric a card prints. The practical advice is simple: read GLM-5.2's scores as a snapshot of June 2026, move to GLM-5.3 for new work, and always rerun two or three candidates on your own tasks. For other families see DeepSeek, Kimi and Qwen.

Product names are used descriptively. We are not affiliated with Z.ai.