GLM-5.3 vs. GLM-5.3 Flash on DeepSWE: Cost, Coding, and Routing

GLM-5.3 vs. GLM-5.3 Flash on DeepSWE: Cost, Coding, and Routing Webflow Analyze/Optimize tracking bridge -->

🚀 DeepSeek V4 Pro 0813 vs. GPT-5.6 Sol on DeepSWE →

🤝 Together AI & Y Combinator announce partnership to deliver the first dedicated YC GPU cluster →

⚡ On-demand B200s now available on Together GPU Clusters →

🚀 Now serving MiniMax-M3 for efficient inference →

All blog posts

Model Library

Published 8/28/2026

GLM-5.3 vs. GLM-5.3 Flash on DeepSWE: Cost, Coding, and Routing

GLM-5.3 wins the first try, Flash nearly matches it by the fourth at 17x lower cost, and the cascade beats both: 80.9% at \$1.70 a task.

Authors

Zain Hasan, Shobhit Dixit

Table of contents

40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production...

Key Takeaways

Run GLM-5.3 Flash first and escalate to the full GLM-5.3 only when your tests reject the answer. That cascade solves 80.9% of DeepSWE tasks at \$1.70 each. GLM-5.3 alone solves 69.0% at \$3.99. Twelve points better, 57% lower cost. If you are picking a single model, pick the Flash: it trails by 5.6 points on the first try but only 2.6 points by the fourth, because what distillation removed was reliability, and reliability is what retries buy back. The gap is a first-try effect. 69.0% vs 63.4% at pass@1, four points at pass@2, and 2.6 points at pass@4 (87.6% vs 85.0%). The price gap is 17x. \$3.99 per rollout vs \$0.24. Per \$100 spent, the Flash solves 264 tasks and the full model solves 17. Capability survived, consistency did not. Of the full model's 99 solved tasks, the Flash still solves 93. It stabilizes 15 tasks the parent was flaky on and cracks 3 the parent walls entirely. The Flash cannot convert effort into a win. On flaky tasks, the longer run is the passing run 61% of the time for the full model and 46% for the Flash, below the coin-flip line. One genuine regression: caution. The Flash breaks an already-passing baseline test in 6.9% of rollouts against the full model's 4.4%, so gate it with a regression run.

Available now · US-hosted

Run GLM 5.3 Flash on Together AI

17× cheaper per rollout, function calling and JSON mode, OpenAI-compatible API, served from US infrastructure.

View the model

‍ GLM-5.3 and GLM-5.3 Flash are the same family at two sizes. GLM-5.3 is the full open-weight model; GLM-5.3 Flash is its distilled sibling at one-seventeenth the rollout price. The question every team asks about a flash variant is how much quality it gives up. On DeepSWE, a benchmark that tests a model's software engineering ability across many task types and programming languages, the answer is smaller than the headline pass@1 gap suggests, because the Flash is not a scaled-down clone of the full model. It is a redistribution of its abilities, and the thing it actually lost is not the thing the headline numbers point to.

DeepSWE · At a Glance

GLM 5.3 vs GLM 5.3 Flash, the headline numbers

Model pass@1 Avg cost Solves per $100 Avg out tok Avg steps

GLM 5.3 [max] 69.0% ± 2.7 $3.99 17 80k 125

GLM 5.3 Flash [max] 63.4% ± 4.1 $0.24 264 73k 123

113 DeepSWE tasks · 4 trials per config · both at max effort · 900 rollouts total

We ran GLM-5.3 (max) against GLM-5.3 Flash (max) on all 113 DeepSWE tasks, four trials each, from the published per-trial records: 900 rollouts in total, 452 full and 448 Flash. What follows is the full comparison and the distillation autopsy in one place. Every figure below comes from this run, so it can differ from other public GLM-5.3 vs GLM-5.3 Flash scorecards.

The DeepSWE scoreboard: pass@1 and pass@k Single shot, the full model leads by 5.6 points, 69.0% to 63.4%. That lead is a first-try artifact. It narrows to four points at pass@2 and to 2.6 points at pass@4, 87.6% against 85.0%. Distillation cost the Flash more of its single-shot polish than its ceiling. For any best-of-k workload the effective quality loss is under three points, at one-seventeenth the price. The collapse from 5.6 to 2.6 is the first clue to what changed, and the next section explains it.

What distillation actually cost: consistency, not capability Classify every task, for each model, as a wall (0 of 4 passing), flaky (some passing), or solid (all passing), then trace where the full model's tasks land in the Flash. The top-right zero carries the section: of the 48 tasks the full model solves four for four, not one becomes unsolvable in the Flash. Half stay perfectly solid; the other half only lose consistency. Coverage retention says the same thing: of the 99 tasks the full model solves at least once, the Flash still solves 93, or 94%. Distillation did not remove the ability to solve these tasks. It removed the reliability. A consistency loss, unlike a capability loss, is exactly what retries recover, which is why the pass@4 gap is so much smaller than the pass@1 gap.

Cost comparison: GLM-5.3 vs. GLM-5.3 Flash pricing and speed At \$0.24 against \$3.99, the Flash costs 17x less per rollout: 264 solves per \$100 to the full model's 17. The usual flash tradeoff, slower per token, does not appear. The Flash is quicker, averaging 26 minutes to the full model's 35, and it does not get there by cutting corners: it runs the same number of steps, 123 against 125. The entire speed difference is per-step latency, 12.5 seconds per step for the Flash versus 17.0 for the full model, a smaller model executing each step about 27% faster in a smaller working window (145k average peak context to the full model's 155k). Lower cost, quicker, thinking the same length.

Where the consistency went: distillation erodes from the margins Group tasks by how reliably the full model solved them, then measure the Flash's mean pass rate in each group. The result is a clean monotonic gradient: the more reliably the full model held a task, the more of that ability the Flash keeps. Distillation eats from the edges of the capability distribution and leaves the core intact. It sharpens the boundary between what the model can and cannot do rather than scaling the whole profile down.

GLM-5.3 Flash lost the ability to take a hard attempt to a win This is the mechanism behind the flakiness, and the most surprising result in the study. On a flaky task, the question is whether the longer run tends to be the winning run. Same task, same difficulty, only the run varies. Full GLM-5.3: on 61% of its flaky tasks, the passing run took more steps than the failing run. The full model...