A benchmark is a repeatable test with a defined task and scoring rule. It can tell you how a model performed under those conditions, but not whether it will solve every task you give it. Price matters only after that distinction is clear. For General Language Model 5.3-Flash (GLM-5.3-Flash), the launch table is mostly self-reported by Z.ai; the figures marked independent come from Artificial Analysis (AA). Use the board below to separate the tests: click intelligence, agentic, speed, and value rather than treating one number as a verdict on the whole model.
intelligence per dollar
The AA Intelligence Index is a composite score, so it summarizes several evaluations without identifying one particular skill. Flash scores 57, versus a ~27 median for open models of the same size. Z.ai's claim is that this places intelligence previously available at roughly 10× the cost at a lower price.
vs the previous tier
On Z.ai's launch table, Flash beats GLM-5.2 on all six listed tests: DeepSWE (63.4 vs 46.2), AutomationBench (48.8 vs 26.2), Terminal-Bench 2.1 (84.3 vs 81.0), and three others, at one-tenth the price. Those tests measure different kinds of work, so their scores are not interchangeable. On Z.ai Code Bench, Flash scores 29.0 at maximum effort versus Opus 4.8's 29.5. In the board's speed tab, rates are tokens per second and time to first token (TTFT) is the wait before output begins; the value tab prices output per one million tokens.
Compare benchmark results
Click a tab, then compare the bars and notes.
Artificial Analysis Intelligence Index v4.1.1, higher scores are better
- GLM-5.3-Flash: same-size open-weight median ≈ 27
- GLM-5.3 (full flagship): Artificial Analysis embedded data via OpenRouter
- Same-size open median: Artificial Analysis class median
Flash's picture is mixed. It scores 57 versus a class median of 27 on intelligence per dollar and beats the previous tier on all six listed Z.ai launch benchmarks at one-tenth the price. It starts faster than its class median but streams more slowly, 49 versus 67 tok/s. “Flash” describes the bill, not the stopwatch. Launch figures are self-reported; independent Artificial Analysis figures are marked above.
What the scoreboard does not settle
- “Flash” labels price, not throughput. Output speed measures how many tokens arrive each second after generation begins. Flash produces about 49 to 50 tokens per second, versus a class median of 67. Its time to first token, the delay before output begins, is about 1.5 seconds and beats the 2.14-second median. Choose it for cost when steady output speed is not the constraint.
- Thinking is always on. There is no non-thinking mode, only low, high, and maximum effort levels. The model's reasoning process still consumes output-token budget at each level, while the listed price stays the same. Plan that budget before comparing a short answer with a long agent run.
- Independent long-context evidence does not exist yet. A long-context test asks a model to use information far back in its input. An ablation is a controlled comparison that changes one design choice; the GLM-5 report's own ablations (§2.1.2) show that linear-attention hybrids retain a small retrieval gap on needle-in-a-haystack tasks, where the model must find a planted detail in a long document. Flash uses 11 lossless sparse layers plus Kimi Delta Attention (KDA), a fixed-state linear-attention method, and bets that the two recall paths cover that gap. Nobody has published results for the shipped model on RULER, the long-context benchmark titled “What's the Real Context Size of Your Long-Context Language Models?”Demand those numbers before relying on one-million-token recall for a legal workload.
- Multimodal input does not mean multimodal output. Multimodal means the model accepts more than one kind of input. Text, images, and video go in; text comes out. The vision tower, a 24-layer Vision Transformer (ViT), is included in the same checkpoint, the saved set of model weights.
The price has a concrete engineering explanation. For each output token, Flash activates 18 billion of its 320 billion parameters, leaving 302 billion inactive; it keeps 34 layers' attention state fixed in megabytes and limits the cache that grows per token; it reads selected context instead of all of it; it checks speculative draft tokens in batches; and it leans on memory bandwidth during decoding. Each choice removes a specific cost, but none turns a launch table into independent proof. Use the board and receipts to separate reported scores, independent measurements, and the long-context questions that still need answers.