跪拜 Guibai
← Back to the summary

GLM-5.3-Flash Matches Claude Opus 4.8 at 1/40th the Cost, Runs on Chinese Silicon

Today, Zhipu AI launched and open-sourced GLM-5.3-Flash (320B-A18B) — the first natively multimodal model in the GLM-5 series.

The official announcement can be condensed into one sentence: it scores 57 on the Artificial Analysis Composite Intelligence Index, matching Claude Opus 4.8; with a limited-time promotional price of 4.5 cents per task, which is 1/40th the cost of Opus 4.8.

To put it another way: the kind of intelligence you previously had to ration at luxury prices is now supplied at residential electricity rates.

Should frontier intelligence be bound to frontier pricing? This is the industry default that Zhipu AI aims to disrupt today.

Let's look at the official announcement first:

Z.ai official X announcement: GLM-5.3-Flash officially released, with six benchmark comparisons

The announcement is brief, but three pieces of background information are worth highlighting:

The speed of dissemination is also part of the context: the announcement was posted at 22:12 that evening, and within an hour and a half, it had 960,000 views and 11,000 likes.

Frontier Intelligence Has Always Been Priced as a Luxury

Over the past two years, an unwritten equation has governed the industry: frontier intelligence = frontier pricing.

Independent developers dared not let agents run overnight — waking up to a bill potentially higher than their rent.

Startup teams treated the decision to 'use a flagship model or not' as an architectural decision on par with fundraising.

Heavy users each had their own money-saving playbook: truncate context, limit turns, downgrade whenever possible.

The postures varied, but the essence was the same: token-metered intelligence had to be used sparingly.

'Niu Lai,' So It Was You

There was an interlude before the release.

Zhipu AI placed GLM-5.3-Flash on OpenCode and OpenRouter under the anonymous model name 'Ox-Alpha' to gather real-world feedback, and the Chinese community gave it a nickname: Niu Lai (The Bull is Coming).

It became the most popular model on both platforms that week, setting new call-volume records on OpenCode and OpenRouter.

Reveal: The anonymous model Ox-Alpha, which set new call-volume records on both platforms that week, is actually GLM-5.3-Flash

In other words, many people had already used it without knowing its identity or price, and their impression was 'this model is both powerful and cheap.'

There was also a detail no one knew at the time: all this traffic was served entirely by domestic chips.

Performance: Going Toe-to-Toe with Opus 4.8

Let's look at the hard data first.

GLM-5.3-Flash: 320B total parameters, 18B activated, 45 layers, the first natively multimodal model in the GLM-5 series.

It scores 57 on the AA Composite Intelligence Index (v4.1.1), entering the global frontier model tier and matching Anthropic's most popular Claude Opus 4.8; at a promotional price of 4.5 cents per task, this level of intelligence previously cost roughly 10 times more.

AA Intelligence Index vs. Cost per Task (log scale): At the same 57-point intelligence level, Opus 4.8's cost per task is approximately 40 times that of GLM-5.3-Flash

Coding and Agent benchmarks comprehensively surpass GLM-5.2, which has double its parameter count: DeepSWE v1.1 63.4 vs 46.2, AutomationBench 48.8 vs 26.2.

Against Opus 4.8, it's a close fight: DeepSWE 63.4 vs 58.0, Toolathlon Verified 78.4 vs 76.2, AutomationBench 48.8 vs 41.0, surpassing it on multiple metrics; in the subjective evaluation of the self-developed Z.ai Code Bench, under max effort, it scores 29.0 vs 29.5, nearly a tie.

Six coding and Agent benchmarks: GLM-5.3-Flash comprehensively surpasses GLM-5.2, and surpasses Opus 4.8 on multiple metrics including DeepSWE, Toolathlon, and AutomationBench

One model, matching the capability of a top-tier flagship, at one-fortieth the price. The price difference in between is not a subsidy; it's the architecture.

Architecture: Built for 1/40th the Price

Three keywords.

Hybrid Attention. The first open-source frontier model to adopt a sparse attention + linear attention hybrid architecture: linear attention captures local dependencies via recurrence, while sparse attention uses a lightweight indexer to recall global context, structurally reducing the service cost of long contexts.

IndexPool. Through weighted pooling, it compresses the indexer's 4 cache vectors into 1, specifically reducing the indexer's latency and memory overhead under a 1M token context.

mHC (Manifold Constrained Hyperconnection). Further improves the model's scaling efficiency.

Let's do the overall math: total parameters are comparable to GLM-4.5 (320B vs 355B), but activated parameters dropped from 32B to 18B, and layers from 92 to 45, nearly halved; combined with the latest 30T token multimodal pre-training corpus.

The result: compared to GLM-5.3, attention computation is reduced by 3.01x, and KV cache by 4.44x; in a comparison among GLM-5.3, DeepSeek-V4-Flash, and Kimi-K3, it has the lowest single-token attention computation.

Comparison of attention computation per head, per layer, and KV cache per layer: GLM-5.3-Flash has the lowest computation among the four, while KV cache is still slightly higher than Kimi-K3 and DeepSeek-V4-Flash

To be restrained, there are shortcomings: the KV cache is still slightly higher than Kimi-K3 and DeepSeek-V4-Flash, which the official team acknowledges as the next optimization target.

Vision Enters the Coding Loop, Coding Enters Professional Work

The first case already exceeds the literal definition of 'programming'.

Without any external assets, GLM-5.3-Flash autonomously ran inside Blender for 16 hours, constructing a professional chef's residence and test kitchen of approximately 400 square meters.

Geometry, materials, lighting, and spatial relationships must remain consistent across all viewpoints — turning world knowledge into a verifiable 3D structure. Coding became the proxy through which the model expresses and validates its own knowledge.

The second case closes the loop down to the pixel.

Inside ZCode, the Browser Use Agent and Computer Use Agent allow the model to coordinate across code, browser, and graphical interface: it writes, renders, views, and revises on its own.

Many problems only become apparent once rendered and interacted with — with native visual capability built-in, the model can judge for itself when it needs to 'take a look.'

Initial output for the same frontend task: layout overflow, a problem only visible after rendering

Final render after visual self-verification: the model discovered the problem itself and fixed it itself

The third case: generation that is deliverable as-is.

Finance: runs the full pipeline from sourced financial research and report generation to modeling analysis, with traceable references.

Legal: reviews contract fees, accounts, and liability clauses, annotating according to lawyer conventions; drafts lawyer's letters, contracts, and litigation documents, with formatting compliant with practice norms.

On the professional work benchmark GDPval-AA v2, it scored 1773 points, the highest among all compared models (including Opus 4.8's 1582 and GPT-5.6 Terra's 1571); on OfficeQA Pro, it scored 62.4, significantly higher than Opus 4.8's 48.9.

An honest footnote: on pure visual perception benchmarks like BabyVision, there is still a gap compared to GPT-5.6 Terra (61.6) and Gemini 3.7 Flash (70.9) — its visual strength lies in 'reasoning and coding with tools,' not bare perception.

Three Things Truly Worth Recording

More worth discussing than the model itself are these three things.

First, running a frontier model on domestic chips has been validated at scale.

Over the past week, all traffic for Ox-Alpha was served by a domestic chip cluster — tens of thousands of domestic accelerator cards with a self-developed high-bandwidth interconnect. To overcome the limitations of single-card compute and memory, Zhipu AI built a dedicated inference engine on top of SGLang: a production-grade EPD (Encode–Prefill–Decode) disaggregated architecture, W8A8 quantization, INT8/FP8/BF16 mixed cache quantization, and Layer Split.

Compared to the initial baseline on the same hardware, end-to-end service performance improved by 3x, achieving hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs.

The easter egg: the development of this inference engine itself was significantly accelerated by an infra agent driven by GLM-5.3 — helping engineers write kernels, diagnose performance bottlenecks, and improve the deployment stack. The model optimizes the system, and the system hosts the model.

Second, the price anchor for frontier intelligence has been reset.

Ten minutes after the main announcement, Zixuan Li from the Z.ai team posted the official pricing: GLM-5.3-Flash is available at 50% off through the official Z.ai API for a two-week period — and explicitly stated that 'this discount also applies to third-party model aggregators.'

Official pricing comparison posted by Zixuan Li: GLM-5.3-Flash's three token unit prices are all halved, while competitors' listed prices are an order of magnitude higher

The numbers that hit the bill: after discount, input $0.075, output $0.25, cached input $0.015. Placed on the same chart, it's even more stark — Claude Sonnet 5 lists output at $10, GPT-5.6 Terra at $12, and the cheapest, Gemini 3.7 Flash, at $7.5; GLM-5.3-Flash, even at its original price (output $0.5), is a fraction of these figures, and the discounted cached input is just one-tenth of the cheapest competitor.

4.5 cents buys task intelligence that previously cost 10 times more. Once the capability-cost curve is pushed to a new position, the pricing of all 'flagship models' will need to be re-explained.

Third, the weights are open-sourced.

The 320B-A18B weights are on HuggingFace under the MIT license, runnable out-of-the-box with SGLang, vLLM, and TokenSpeed. A 1/40th price anchor, plus an open-source license with virtually no barrier to use, leaves little buffer room for API pricing.

How to Get Started

The official blog's conclusion is fitting to end with: Frontier intelligence need not be accompanied by frontier cost — this is not a single trick, but the result of three-layer synergy among architecture, corpus, and infrastructure, and this recipe is being applied to even larger models.

If you are still rationing flagship models by the token, consider adding this open-source model to your default options and try it for a week.

References

All charts in this article are screenshots from the above official release materials.