Twenty-two benchmark rows sit on the GLM 5.3 Flash card, and two of them are the same run recorded twice under different names [OURS]. That duplication is the most useful thing on the page, because it is a controlled comparison of the two publishers whose numbers a reader is asked to weigh against each other. The glm 5.3 flash pricing query lands on this table. It is served by OrcaRouter.
One benchmark, two labels, three hundredths
| Row as printed | Score | Source |
| Terminal-Bench 2.1 | 84.3 | z.ai |
| terminalbench_v2_1 | 84.27 | artificialanalysis.ai |
The difference is 0.03. That is not two measurements of a model; it is one measurement reported by two parties, and the near-identity says the harness is the same and the run is the same.
Which then raises the question the rest of the table asks. If these two publishers agree to three hundredths on the benchmark they both run, why do their other rows look like different models?
The thirteen vendor rows
Thirteen rows carry `z.ai` as their source, and they cluster into a shape:
| Row | Score |
| CharXiv Reasoning (with tools) | 89.4 |
| MMVU | 80.5 |
| Chartography (with tools) | 78.0 |
| Toolathlon Verified | 78.4 |
| MVBench | 77.8 |
| DeepSWE v1.1 | 63.4 |
| OfficeQA Pro | 62.4 |
| NL2Repo | 56.3 |
| HLE (with tools) | 55.3 |
| BabyVision | 53.4 |
| AutomationBench v1.0.6 | 48.8 |
| Agents’ Last Exam | 26.3 |
This is the multimodal half of the model. CharXiv Reasoning at 89.4, MMVU at 80.5 and MVBench at 77.8 are chart, video and vision benchmarks — and the card lists `text`, `image` and `video` as accepted inputs [OURS]. Z.ai is documenting what the model does with a picture, because most of the competitive set cannot do it at all.
The two low rows are the interesting ones. Agents’ Last Exam at 26.3 and AutomationBench at 48.8 are the tasks that are not yet solved, and a vendor publishing them is publishing where the model stops.
The nine evaluator rows
The independent half is text and code, and it behaves like a mature model’s table:
| Row | Score |
| GPQA Diamond | 91.2 |
| terminalbench_v2_1 | 84.27 |
| Long-Context Recall | 80.0 |
| AA Coding | 71.5 |
| SciCode | 51.6 |
| tau_banking | 47.2 |
| AA Intelligence | 41.8 |
| Humanity’s Last Exam | 39.9 |
| terminalbench_v4_0 | 32.8 |
The pattern here is the familiar one: knowledge saturated in the nineties, coding in the low seventies, and the newest hard generation — v4.0 — at 32.8, which is where most of the catalogue lands.
The tools delta is the row worth computing
The table prints Humanity’s Last Exam twice: 39.9 without tools on the evaluator’s harness and 55.3 with tools on the vendor’s [INDEPENDENT and VENDOR].
Those are not the same benchmark configuration, so the difference is not pure tool gain — but 15.4 points is large enough to say something. A model that gains a third of its HLE score when it can call a tool is a model whose knowledge is weaker than its ability to go and find things.
That reading fits the rest of the card. Long-Context Recall at 80 and the multimodal rows in the eighties describe a model good at handling material it is given; HLE at 39.9 describes one that struggles with material it has to know.
For a workload, that is a routing rule: give this model the document, the screenshot, the table. Do not ask it what it remembers.
Why the duplicate row matters for a buyer
A reader comparing cards sees GLM 5.3 Flash at 84.3 on Terminal-Bench and Claude Fable 5 at 84.64 on terminalbench_v2_1, and has no way to know whether those are the same benchmark from the same publisher.
They are. The Fable card labels its row `terminalbench_v2_1` and tags it artificialanalysis.ai; the GLM card prints both that row and the vendor’s own Terminal-Bench 2.1 label at 84.3. Same test, same ballpark, two names.
That is the practical argument for printing both rows even when they agree: it is the only way a cross-card comparison can be made without knowing which publisher a competitor’s number came from.
The takeaway
GLM 5.3 Flash carries 22 benchmark rows, thirteen from Z.ai and nine from the independent evaluator, and one of them appears twice at 84.3 and 84.27 [OURS].
The duplicated row is the card being honest about its own sources. The rest of the table is a multimodal model with a weak recall of things it was not given: strong on charts, video and long documents, 39.9 on Humanity’s Last Exam until you let it use a tool.