GLM 5.3 Flash vs DeepSeek V4.1 Flash: Price, Hardware, and Benchmarks (2026)

DeepSeek V4.1 Flash landed on September 10 and changed the answer to “GLM 5.3 Flash or DeepSeek Flash?” Here’s the new comparison, two weeks after the old one.

Two weeks ago the GLM 5.3 Flash vs DeepSeek V4 Flash call was easy to summarize: GLM scored higher and cost less on the API, DeepSeek was faster and fit on two GPUs. Then DeepSeek retired V4 Flash and shipped V4.1 Flash: 552B parameters, a new encoder-decoder architecture, native image input, a lower price, and a checkpoint three times the size of the one it replaced. Most of the old comparison no longer holds. This is the current one.

The two models side by side

Three rows changed the comparison: the checkpoint size, the price, and the modalities. Take them in order.

Hardware: the two-GPU advantage is gone

The strongest argument for DeepSeek’s Flash used to be hardware. V4 Flash shipped pre-compressed at 166.9 GB and ran on 2x H200; GLM 5.3 Flash’s 306 GiB checkpoint needed a full 8-GPU node. That was a two-GPU-pod-versus-one-node difference every hour the model was up, and it decided most self-hosting conversations before benchmarks came up.

V4.1 Flash is roughly 510 GB of FP8 safetensors. That’s more than four H200s hold before overhead, so the realistic floor is the same 8-GPU node GLM needs, and it’s the tighter fit of the two: on 8x H100 80 GB (640 GB), GLM leaves about 300 GB for KV cache and batch, V4.1 Flash leaves about 130 GB. On 8x H200 or 8x B200/B300 both run comfortably. The KV cache picture partly offsets that: DeepSeek’s second-generation compressed sparse attention brings V4.1 Flash’s persistent cache to 890 bytes per token, so a full 1M-token context costs well under a gigabyte, while GLM’s hybrid sparse-plus-linear attention is a long-context memory win in its own right but not in that class. In practice, GLM has more room for concurrent requests on the same node; V4.1 Flash has more room for very long contexts. The DeepSeek V4.1 Flash hardware requirements post has the verified node configs and the one option that might bring 4x H200 back into play.

The bottom line on hardware: it’s no longer a reason to pick DeepSeek. Both are node-scale models. If you were self-hosting V4 Flash on two H200s for the price, that deployment still works with the V4 Flash weights, which remain available under MIT. It just isn’t the current model anymore.

Price: closer than it looks, and it depends on the clock

GLM 5.3 Flash is $0.15 per million input tokens and $0.50 output, flat, all day. DeepSeek V4.1 Flash is $0.15 and $0.60 off-peak and $0.30 and $1.20 at peak, where peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, roughly Beijing business hours.

So for a US team running daytime traffic, the two models are nearly the same price: identical on input, GLM 10 cents cheaper per million on output. During DeepSeek’s peak window, GLM is half the price on input and less than half on output. Across a 24-hour workload the average lands in between and depends on your traffic curve.

Caching flips it. DeepSeek charges $0.003 per million for cache hits off-peak, GLM charges $0.03, a 10x difference. An agent loop that resends a 20K-token system prompt and tool schema every turn is mostly cache hits, and on that traffic DeepSeek is far cheaper as long as you’re outside the peak window. A workload with little reuse, short prompts and long generations, favors GLM’s flat $0.50 output rate.

Neither is on Yotta AI Gateway yet, where the flat per-token rates with no peak windows would remove the clock from the calculation; the Gateway carries DeepSeek V4 Flash and V4 Pro and the GLM 5.3 flagship today, and this post will note it when either Flash joins.

Benchmarks: one overlap, two vendors, no referee

Independent numbers arrived on September 14. Artificial Analysis scores DeepSeek V4.1 Flash at 40 on its Intelligence Index (version 4.3) against 42 for GLM 5.3 Flash and 35 for the retired V4 Flash. Two points is inside the noise on that index, so on the composite these are peers. The sub-scores split the way the vendors' own tables suggested: V4.1 Flash leads on AutomationBench (69% vs 60%) and the long-context test AA-LCR (84% vs 80%); GLM 5.3 Flash leads on Terminal-Bench 4.0 (33% vs 27%); they tie on Humanity's Last Exam (40% each) and SciCode (52% each).

The single benchmark that appears in both vendors’ launch tables is DeepSWE v1.1. DeepSeek reports 74.2 for V4.1 Flash at maximum reasoning effort; Z.ai reports 63.4 for GLM 5.3 Flash. Those are two labs’ runs of the same test under their own settings, so treat the 11-point spread as a signal, not a measurement. It's consistent with DeepSeek's stated focus for V4.1 Flash, agentic coding and terminal work, where its own table also shows it passing V4 Pro and GLM's flagship GLM 5.3 (66.9 on DeepSWE, per Z.ai); the GLM 5.3 vs DeepSeek V4.1 Flash comparison takes that flagship-versus-Flash question on directly. On knowledge-heavy tasks without tools, V4.1 Flash’s own numbers are Flash-class (HLE 36.8, GPQA Diamond 90.9), and GLM 5.3 Flash’s table doesn’t report those, so there’s nothing to compare.

Speed is not a tie. Artificial Analysis measures V4.1 Flash at 214 output tokens per second on DeepSeek's API with a 1.17-second time to first token, against 107 tokens per second and 2.45 seconds for GLM 5.3 Flash on Z.ai's. That's DSpark speculative decoding, built into V4.1 Flash, doing what it did for V4 Flash. For latency-sensitive agent loops, a 2x decode advantage is the biggest single difference between these two models.

Modalities, licenses, and the rest

Both models now take images. That was GLM’s exclusive two weeks ago; V4.1 Flash’s from-scratch vision encoder closes it. GLM 5.3 Flash still takes video input, which V4.1 Flash does not, so if video is in the workload GLM is the only option of the two.

Both are MIT. Both list day-one vLLM and SGLang support. GLM 5.3 Flash has two extra weeks of serving-stack maturity and community configs behind it; V4.1 Flash has a brand-new architecture that the engines are still optimizing for. If you hit an odd engine bug in the first month, it’s more likely to be on the DeepSeek side.

Which one to use

Choose GLM 5.3 Flash if you want flat, predictable API pricing without watching a clock, if video input is in the workload, if you'll self-host and want the extra headroom on an 8x H100 node, or if you want the model with a month of production mileage behind it.

Choose DeepSeek V4.1 Flash if your traffic is agentic and cache-heavy and mostly off-peak, where the $0.003 cache-hit rate makes it the cheapest frontier-adjacent model available, if decode speed matters (it's twice as fast on the hosted APIs), or if very long contexts matter and the 890-byte KV cache pays off.

If you’re already on V4 Flash, the question is different: keep the two-H200 deployment on the V4 Flash weights, or move to the API. The upgrade path to V4.1 Flash is a node, not a pod.

Both speak OpenAI-compatible interfaces, so running the bake-off on your own traffic is a base URL and a model string, and that’s the only benchmark here that isn’t a vendor’s.

How to run each today

DeepSeek V4.1 Flash: DeepSeek’s API as deepseek-flash, or the weights from Hugging Face on an 8-GPU H200 or B300 node by the hour. The V4.1 Flash launch post has the full pricing table, what happened to the V4 Pro routing plan, and the hardware estimates.

GLM 5.3 Flash: Z.ai’s API at $0.15 / $0.50, or self-hosted on an 8-GPU node using the serving stack our GLM 5.2 SGLang guide walks through. The GLM 5.3 Flash hardware breakdown covers the VRAM math.

For a three-way test against the models that are on the Gateway now, DeepSeek V4 Pro and GLM 5.3 are both live on Yotta AI Gateway with flat pricing, one key, and the same OpenAI-compatible interface.

Frequently asked questions

Is DeepSeek V4.1 Flash better than GLM 5.3 Flash? On the composite, they're level: Artificial Analysis scores GLM 5.3 Flash 42 and V4.1 Flash 40 on its Intelligence Index. V4.1 Flash is twice as fast on the hosted APIs (214 vs 107 tokens per second) and leads on AutomationBench; GLM 5.3 Flash leads on Terminal-Bench 4.0. The vendors' own DeepSWE runs (74.2 vs 63.4) favor DeepSeek but aren't a controlled comparison.

Which is cheaper, DeepSeek V4.1 Flash or GLM 5.3 Flash? Off-peak they’re almost the same: $0.15 in for both, $0.60 out for DeepSeek versus $0.50 for GLM. At DeepSeek’s peak (weekday Beijing business hours) GLM is half the price or better. For cache-heavy workloads DeepSeek wins off-peak, since its cache-hit rate is $0.003 against GLM’s $0.03.

Which needs less hardware? GLM 5.3 Flash, now. Its FP8 checkpoint is about 306 GiB against roughly 510 GB for V4.1 Flash. Both need an 8-GPU node; GLM leaves more headroom on 8x H100. The two-H200 option that DeepSeek V4 Flash offered is gone with V4.1 Flash.

Which handles images and video? Both take image input. Only GLM 5.3 Flash takes video input.

Are they open source? Yes, both ship full weights on Hugging Face under MIT.

What happened to DeepSeek V4 Flash? DeepSeek retired it from the API on September 10, 2026; the deepseek-v4-flash name now serves V4.1 Flash. The V4 Flash weights are still on Hugging Face and still run on two H200s. The original GLM 5.3 Flash vs DeepSeek V4 Flash comparison remains accurate for those weights.

Are they on Yotta AI Gateway? Not yet. Yotta AI Gateway carries DeepSeek V4 Flash and V4 Pro and the GLM 5.3 flagship at flat rates. This post will be updated when V4.1 Flash or GLM 5.3 Flash is added.

Bottom line

V4.1 Flash took away DeepSeek’s hardware advantage, matched GLM on input price, nearly matched it on output, and closed the vision difference, while adding a cache-hit rate ten times cheaper and a set of coding benchmarks that, if they hold up independently, put it ahead. GLM 5.3 Flash keeps flat pricing, video input, a smaller checkpoint, and a two-point edge on the one independent score, which is a tie in practice.

Two weeks ago these were two different products sharing a price tier. Now they’re direct competitors, and the deciding factor is your traffic: cache-heavy and off-peak leans DeepSeek, flat and predictable leans GLM. Run both against your own eval before the independent numbers arrive, and keep the model name in config so you can switch when they do.