GLM-5.3-FlashX vs DeepSeek V4.1 Flash: The Speed Tier That Is Slower Than the Cheap Model

Z.ai's premium speed tier has a problem, and it is called DeepSeek V4.1 Flash. When Z.ai launched GLM-5.3-FlashX on September 18, 2026, it sold the new tier on one number: up to 200 output tokens per second, roughly five times the throughput of the GLM-5.3-Flash it is built from, for 2.5× the price. Independent measurement already puts DeepSeek's budget model at 214.7 tokens per second with a 1.15-second time to first token — faster on both counts than the ceiling Z.ai is advertising, and at a price FlashX does not come close to. If you are buying speed, the market moved before FlashX arrived.

That framing is unfair to FlashX in one specific way, and fair in every other. Both models are 1M-context open-weight derivatives that were engineered for cost, and both are genuinely good at what they were built for. But they were built for different halves of the same problem, and the price gap between them is now wide enough that "which is better" has a one-line answer for most workloads.

Where each one is actually ahead

• Price, list — GLM-5.3-FlashX $0.37 / $1.25 per 1M input/output vs DeepSeek V4.1 Flash $0.15 / $0.60 off-peak, $0.30 / $1.20 at peakbr> • Cached input — FlashX $0.075 per 1M vs DeepSeek $0.003 off-peak, a 25× gap that compounds hard in agent loopsbr> • Measured throughput — FlashX up to 200 tok/s (vendor peak, no independent figure published) vs DeepSeek 214.7 tok/s (Artificial Analysis, on DeepSeek's API)br> • Time to first token — not published for FlashX vs 1.15s measured for DeepSeek V4.1 Flashbr> • Independent intelligence — FlashX inherits GLM-5.3-Flash's 42 on the Artificial Analysis Intelligence Index vs DeepSeek V4.1 Flash at 40br> • Architecture — 320B total / 18B active MoE vs 552B total with roughly 8B active on prefill and 16B on decodebr> • Input modality — text, image, video and files vs text and imagebr> • Output cap — 131,072 tokens vs roughly 384,000br> • Open weights — GLM-5.3-Flash is MIT-licensed; FlashX is a hosted tier with no weights of its own, and DeepSeek V4.1 Flash is MIT-licensed

Read that list twice, because the shape of it is the story. FlashX wins exactly two rows, and both are narrow: a two-point edge on an independent intelligence index, and video input. DeepSeek wins price, cached price, measured speed, output length and modality breadth in the sense that matters — it takes images and produces long answers. The two-point index gap is within the noise of a single benchmark run and should not be the thing that decides your architecture.

The speed tier that is slower than the cheap model

This is the part worth dwelling on. Z.ai built FlashX to solve a real problem: GLM-5.3-Flash is slow. Artificial Analysis measured it at 98 output tokens per second, which is above the peer median for open-weight models of its size but nowhere near interactive. The company's fix was a serving-stack rebuild — roughly 100,000 domestic accelerators, an SGLang-based engine, W8A8 quantization, mixed-precision KV cache, and an encode–prefill–decode disaggregated architecture — and it reports about 3× end-to-end throughput over its baseline on the same hardware.

DeepSeek solved the same problem at the architecture layer, and got there first. V4.1 Flash's Causal Encoder–Decoder split activates about 8B parameters on prefill and 16B on decode rather than a flat figure, and Compressed Sparse Attention 2 with FP4 KV caching cuts the global cache to 890 bytes per token — roughly a quarter of the HBM and an eighth of the SSD storage its predecessor needed. The result is a model that runs at 214.7 tokens per second as measured by a neutral lab, on the same first-party API, with no premium tier required.

Z.ai's 200 is a vendor peak and DeepSeek's 214.7 is an independent median, which is the least favourable possible comparison for Z.ai and the reason to hold it loosely. It is entirely possible FlashX beats DeepSeek on a warm, short-output, uncongested request. It is not possible to know, because nobody outside Z.ai has published FlashX throughput numbers. What is knowable today is that the two models are in the same speed band, and one of them costs a third as much.

Where DeepSeek's price advantage becomes structural

DeepSeek V4.1 Flash charges $0.15 per million input tokens and $0.60 per million output tokens off-peak, doubling to $0.30 and $1.20 during two weekday windows (01:00–04:00 and 06:00–10:00 UTC). FlashX is $0.37 and $1.25 flat. Line those up and the flat pricing that looks like a feature is actually the more expensive structure for anyone who can schedule work.

The cached-input line is the one that decides agent workloads. DeepSeek charges $0.003 per million cached input tokens off-peak; FlashX charges $0.075, twenty-five times more. An agent that re-sends a long system prompt and tool schema on every step pays that difference on every step, and it is the single line item most likely to dominate a real bill. Z.ai lists cache storage as free for a limited time, which is a discount on storage rather than on the reads that actually accumulate.

There is a counterweight, and it is honesty about what DeepSeek's own numbers cost. The vendor-run evaluations behind V4.1 Flash's headline scores were produced at maximum reasoning effort, and DeepSeek documents that moving from effort 25 to effort 100 lifts DeepSWE v1.1 from 66.0 to 74.2 while consuming roughly 2.5× the output tokens. At 2.5× the tokens, the effective cost of the high-effort configuration is a lot closer to FlashX than the sticker suggests — and the model is documented as very verbose, generating 250M output tokens on the Intelligence Index against a median of 130M. A cheap per-token rate on a chatty model is not the same as a cheap task. Anyone comparing these two on price should compare cost per completed task, not cost per million tokens, and should expect to measure rather than estimate.

How to decide, concretely

If your workload is a long-running agent with a stable prefix, DeepSeek V4.1 Flash is the pick, and the cached-input ratio alone decides it. If you need video understanding or file input in the same call, FlashX is the only one of the two that takes them — that is a genuine capability difference, not a spec-sheet one. If you need long generated outputs, DeepSeek's roughly 384K output cap against FlashX's 131,072 matters for report generation, long code files and anything that synthesises rather than answers. If you need the strongest independent score on a single benchmark index, FlashX's 42 versus 40 is real but too small to build on.

Both models are reachable from one endpoint if your stack is already routed, which is the cheapest way to settle this. On OrcaRouter the vendor list price is passed through at 0% markup, so DeepSeek's off-peak discount and any future Z.ai price move land the same day they happen, and switching between the two is a model string rather than an integration. Automatic failover is worth having for FlashX specifically: it is a three-week-old serving tier on a new cluster, and the failure mode you are insuring against is a capacity ceiling, not a model that stops working.

The blunt summary: DeepSeek V4.1 Flash is the better default for most production work — cheaper per token, cheaper per cached token, measured faster, and capable of longer outputs. GLM-5.3-FlashX is the right call in a narrower band where you need video or file input and want a marginally higher independent index behind it. Z.ai priced a speed tier at a premium and shipped it into a market where the fast cheap model already existed. That is the whole comparison.