GLM-5.3 Flash Ties Grok 4.7 for a Third of the Price, When It Answers

On reference-preserving analytical writing, the most heavily measured long-form task in the Fronset benchmark, three models tie at the top. The cheapest of them, GLM-5.3 Flash, costs a third of what Grok 4.7 does per run, but its score comes from only 52% of attempts [^1].

That moves the decision away from quality, which is tied, and onto whether a call returns an answer at all, which the score does not show. The task is one component of long-form work: no result in the benchmark measures a whole document generated end to end.

Three models tie at the top

The intervals of the top three overlap, so on this evidence they are tied, and they are the only models to reach 9.0 on the task [^1]. Small gaps are normal here. The top two answers finished within half a point of each other in 88.4% of 69 judged occasions, with a median gap of 0.1 [^1].

Price does not buy position either. Claude Opus 5.5 costs the most per run of the four and has the lowest point score. Its interval overlaps Kimi K3’s, so it is not clearly behind that model, but its premium buys no measured quality advantage on this task [^1].

The catch is the answer rate

Every score covers only the calls that returned. GLM-5.3 Flash’s 9.24 describes the attempts that produced an answer, and the rest are not in the number at all. A score means little until you read it beside the answer rate.

Answer rates also vary widely between models. Thinking Machines Inkling scores 8.42 and answers on 90% of attempts, while GLM-5.3, a different model from GLM-5.3 Flash, scores 8.88 and answers on 50% [^1]. A higher score can come with a much lower chance of getting anything back.

The benchmark does not turn answer rate into an effective cost, so this evidence cannot say how much of GLM-5.3 Flash’s price advantage survives once retries are counted. What it does say is that an answer rate this low argues for retries and a fallback [^1].

How to route analytical writing

Send the work to GLM-5.3 Flash first, retry when it does not answer, and fall back to Grok 4.7 [^1]. There is no measured reason to pay for Claude Opus 5.5 on this task.

Plan spend beyond the per-run price. On this task, 5.9% of spend was lost to output that could not be parsed [^1]. The findings do not attribute that loss to any model, so it cannot rank them, but it belongs in the budget whichever model you choose.

The caveat that matters most is that the comparison is one-sided. The evidence quotes an answer rate for GLM-5.3 Flash but none for Grok 4.7 or Kimi K3. It shows the cheap model’s weakness without showing that the alternatives avoid it, so measure your fallback’s answer rate on your own traffic before relying on it.

The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your writing needs and compare the prices of the models that clear it. Check these figures there, and test your shortlist and its fallback on your own material, before committing a model to a production path.