GLM-5.3-Flash vs GLM-5.3: Four Points for Nine Times the Price

Zhipu released GLM-5.3-Flash on 26 August, twelve days after GLM-5.3. Both are open weights, both hold a million tokens of context, and they reported four of the same benchmarks, which makes this one of the cleanest model comparisons available right now.

Read that chart carefully, because it does not say what a "Flash" model usually says.

What the four numbers show

On the two coding benchmarks Flash gives up under four points. On AutomationBench it is fractionally ahead. The one real gap is Humanity's Last Exam with tools, where the larger model is seven points better, and that is the benchmark of deep, long-chain reasoning, which is exactly where you would expect 753 billion parameters to beat 320 billion with 18 billion active.

That pattern is the useful finding: the smaller model holds its own on doing things, and gives ground on thinking hard about things.

The price

$0.15 and $0.50 against $1.40 and $4.40. Flash is about nine times cheaper on both sides.

Put that beside the table above and the trade is stark: roughly 4% of the coding capability for roughly 89% of the cost.

What Flash adds

This is the part that makes it more than a discount model.

Native multimodality. GLM-5.3-Flash takes text, images and video. GLM-5.3 is text-only. Within one family, the cheaper model is the one that can read a screenshot, a diagram or a scanned invoice, which reverses the usual assumption that vision costs extra.

An MIT licence you can act on today. Flash's weights are published under MIT at 320B with 18B active, around 306 GiB for the FP8 checkpoint. GLM-5.3 was announced as open-weights with the release to follow after safety evaluation, so Flash is the one to plan self-hosting around right now.

Hybrid attention. Flash combines linear-attention layers with sparse attention, which is how the long-context serving cost comes down while a 1M window is preserved. It routes 8 of 288 experts per token.

The specifications

Which to use

Start with Flash. For the routine majority of engineering work — the dependency bump, the endpoint that follows the pattern of the four beside it, the test that needs writing, the log that needs reading — a four-point difference on a vendor-run benchmark will not be visible, and the nine-fold price difference will be.

At the cost-per-solved-task arithmetic, Flash comes out around $0.06 per finished Terminal-Bench 2.1 task against GLM-5.3's $0.55. Same family, same vendor, same weights licence philosophy, nine times the throughput for your budget.

Escalate to GLM-5.3 when reasoning depth is the job. The seven-point HLE gap is the signal to trust here. A migration nobody has attempted, a bug spanning four services, an architectural question with long chains of inference: that is where the larger model earns $1.40.

Escalate on a failure, not a guess. The cleanest rule is to let Flash try first and promote the task when a verifier rejects the result. That way the expensive model only sees work that has demonstrated it needs the money, which is the pattern in running open models at frontier level.

The caveats

Both sets of numbers are vendor-reported by Zhipu, run on Zhipu's harnesses. A vendor comparing its own two models has every reason to be internally consistent, which makes this comparison more trustworthy than most cross-vendor ones, and it is still one vendor's account of its own work.

The AutomationBench figures carry a version note: GLM-5.3's is labelled v1.0.6. Treat the 0.6-point difference as a tie rather than a win.

And as always the number that decides your case is the one from your own repository. Twenty closed tickets through both models will tell you more than this page, and it will tell you specifically whether your work sits in the 4% where the difference shows.

Sources

• GLM-5.3 model card
• Z.ai GLM-5.3-Flash documentation
• Z.ai releases GLM-5.3-Flash