GLM-5.3 scores 60 on the Intelligence Index, but its 170M-token verbosity is the real story

Z.ai’s GLM-5.3 (max) landed this week with a #8 ranking on the Artificial Analysis Intelligence Index, scoring 60 out of a possible 100 against 182 comparable proprietary models. That places it firmly in the frontier tier, above the median of 35 and within striking distance of the top five. But the more interesting number on that page is not the intelligence score. It is the 170 million output tokens the model burned to complete the evaluation suite, against a median of 72 million.

That is 2.4 times the median verbosity. Artificial Analysis flags it plainly: GLM-5.3 is “very verbose.” The model ranks #73 of 182 on the verbosity metric, and its 4-out-of-4 units on that scale reflect volume, not quality. For a reasoning model priced at $1.40 per million input tokens and $4.40 per million output tokens, that verbosity is not a quirk. It is the dominant variable in the cost equation.

The intelligence score is real, but read the fine print

The Intelligence Index v4.1.1 that produced the 60 score blends nine evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. These are not toy benchmarks. Terminal-Bench v2.1 tests agentic coding and terminal use. τ³-Banking measures agentic tool use in a financial context. GDPval-AA v2 scores real-world agentic work tasks. Humanity’s Last Exam and GPQA Diamond probe reasoning and scientific knowledge at the graduate level.

Scoring 60 on that mix puts GLM-5.3 in the same conversation as models from Google, Anthropic, and OpenAI, the other labs whose logos appear on the Intelligence Index versus cost-per-task scatter plot. The model also carries a 1M token context window, roughly 1,500 A4 pages of 12-point Arial, which puts it in the top tier for long-context RAG workloads. On raw capability, this is a serious release.

But the evaluation cost tells a different story. Artificial Analysis spent $1,238.50 to run the full Intelligence Index on GLM-5.3. At $0.68 per Intelligence Index task, the model is “reasonably priced when comparing to other models of similar price,” per the site’s own summary. The median cost per task across the class is higher. The problem is that the per-task number is computed after the token burn. Every extra thousand tokens of reasoning output costs $4.40 at list price. A model that thinks out loud for 170M tokens is a model that bills you for its thinking.

Verbosity is the hidden tax on reasoning models

The reasoning-vs-verbosity tradeoff is not new. OpenAI’s o1 series and DeepSeek’s R1 both generate long chain-of-thought traces. But the magnitude here is notable. The median model in this class produces 72M output tokens across the evaluation suite. GLM-5.3 produces 170M. That gap is not a rounding error; it is a 136% premium on output volume.

For a developer building on GLM-5.3, the math is straightforward. At $4.40 per million output tokens, a task that generates 10,000 output tokens costs $0.044 in output alone. A comparable model generating 4,000 tokens for the same task costs $0.0176 at the same price point. The intelligence score is identical. The bill is not.

This is where the “cost per Intelligence Index task” metric deserves scrutiny. Artificial Analysis weights each evaluation’s cost by its contribution to the index, then divides by task count. That methodology rewards models that are cheap per unit of intelligence. But it does not capture what a production workload actually pays, because production workloads do not run a weighted benchmark suite. They run specific tasks, and those tasks inherit the model’s verbosity profile.

GLM-5.3’s cache discount of 81% mitigates some of this for repeated prompt prefixes. Cache-hit pricing drops to a fraction of the $1.40 input rate, and long-context workloads with stable system prompts will benefit. But cache discounts do not apply to output tokens, and output is where GLM-5.3 is expensive. The 81% cache discount is a salve, not a cure.

What this means for the pricing war

The broader context is that the AI model market has entered a price-per-intelligence arms race. DeepSeek’s open-weights releases forced proprietary labs to cut prices. Alibaba’s Qwen family and Kimi’s models from Moonshot AI have pushed the cost curve down. GLM-5.3 enters that market at $1.40 in and $4.40 out, which Artificial Analysis rates as “moderately priced” against medians of $1.75 and $10.00 respectively. On paper, it undercuts the median output price by more than half.

But the effective price is what the model actually charges per completed task. A verbose model at $4.40 per million output tokens can end up more expensive per task than a terse model at $10.00 per million, if the terse model produces a quarter of the tokens. The median output token count of 72M versus GLM-5.3’s 170M suggests exactly that dynamic is in play. The listed price is the sticker. The verbosity-adjusted price is the real cost.

Artificial Analysis’s own scatter plot of Intelligence Index versus cost per task shows GLM-5.3 sitting near the Pareto frontier. That is a genuine achievement. The model delivers frontier-level intelligence at a competitive per-task cost. But the frontier is defined by the evaluation suite, and the evaluation suite rewards correct answers regardless of how many tokens it takes to produce them. A production system that calls GLM-5.3 for every task will not see the benchmark’s weighted average. It will see the raw token count.

The takeaway for builders

For teams evaluating GLM-5.3, the practical advice is to benchmark on your own workload before committing. The Intelligence Index score of 60 is a capability signal, not a cost signal. Run your actual prompts, measure your actual output token counts, and compute your actual per-task cost. If your tasks are short and require concise answers, a less verbose model at a higher list price may still win on total spend.

There is also a product opportunity here. The market has room for a “verbosity dial” that lets users trade reasoning depth for token cost. GLM-5.3’s non-reasoning variant, which Artificial Analysis notes may exist, could serve that role. But until providers expose that control at the API level, developers are stuck paying for the model’s default behavior.

The deeper point is that intelligence benchmarks are converging. Scores of 60 on the Artificial Analysis Index are becoming table stakes for frontier models. The differentiator is no longer raw capability; it is efficiency. GLM-5.3 proves a Chinese lab can match the frontier on intelligence. The 170M token burn proves that matching the frontier on cost is a separate problem, and one that no benchmark score can hide for long.