Frontline Hotspot
Frontline Hotspot

The Mystery Model "Ox-Alpha" Revealed: Zhipu Open-Sources GLM-5.3-Flash - Claude Opus 4.8 Parity at 1/40th the Price

The anonymous model Ox-Alpha landed on OpenRouter on August 20, topping first-day traffic and ending DeepSeek's 56-day streak; on the evening of August 26, Zhipu revealed it: GLM-5.3-Flash (320B-A18B), now open-sourced. The first natively multimodal GLM-5 model (video/image/text/file inputs), its sparse + linear attention hybrid cuts attention compute 3x and shrinks KV cache 4.4x. It scores 57 on Artificial Analysis' intelligence index, level with Claude Opus 4.8, at roughly 1/40th the price (0.8 input / 2.8 output RMB per million tokens), with all traffic served by 100,000 domestic AI chips and MIT-licensed weights. One twist to remember: cache hits at 0.23 RMB actually cost more than DeepSeek V4-Flash's off-peak 0.05 - coding agents with 90%+ cache-hit rates should run the numbers before switching.

Published August 27, 20266 min read
<!-- zhipu-glm-5-3-flash-open-source-hotspot | hotspot | The Mystery Model "Ox-Alpha" Revealed: Zhipu Open-Sources GLM-5.3-Flash - Claude Opus 4.8 Parity at 1/40th the Price -->

For the past week, overseas developer communities have been dominated by an anonymous model: Ox-Alpha, which landed on OpenRouter on August 20 under a "hidden model" identity and topped the platform's daily call volume on day one, setting an all-time single-day record. Chinese communities gave it a more down-to-earth nickname - "the ox has arrived." On the evening of August 26, the answer was revealed: Zhipu officially launched and open-sourced GLM-5.3-Flash (320B-A18B), confirming that it is Ox-Alpha. It scores 57 on Artificial Analysis's aggregate intelligence index - level with Claude Opus 4.8 - at roughly one-fortieth of the price, all running on 100,000 domestic AI chips.

Scope note: this article is based on reporting by Cailianshe/STCN (reporter Li Mingming), ifanr, and Lanjing News from 2026-08-26/27, plus Zhipu's official documentation and model card. Prices are snapshots; the official pages prevail. Not investment advice.

1. Who Is "the Ox": The Hidden Model That Ended DeepSeek's 56-Day Streak

First, the timeline. On August 20, Ox-Alpha went live on OpenRouter with no attribution: it topped daily call volume on day one and set the platform's all-time single-day record, ending DeepSeek's 56-day reign atop OpenCode. It processed 62T tokens in a single day, became the most popular model of the week on both OpenRouter and OpenCode, and set historical records on both platforms. Over the entire 5-day test window, it absorbed roughly 50T tokens of traffic.

The community got to work. Independent researchers' technical fingerprinting showed that its tokenizer produced token counts identical to GLM-5.3's across 25 test prompts, and that all 11 cross-probe tests matched the GLM family - leading to a "99% confident" attribution to Zhipu. On the evening of August 26, Zhipu announced GLM-5.3-Flash and officially confirmed it was Ox-Alpha. The guessing game ended, but the anonymous-marketing stunt had already done its job: let global developers vote with their traffic first, then reveal the name.

The playbook deserves a closer look. A conventional launch goes "present the parameters, then wait for users to validate them"; Ox-Alpha inverted the order. The model took on real traffic anonymously first - 62T tokens in a day, a 56-day reign ended, historical records on two platforms, all a report card written by users themselves. By the time the community had guessed the answer to 99% confidence from tokenizer fingerprints, the official announcement only had to supply the last 1%. The anonymity itself became the best distribution mechanism: the more people playing the guessing game, the higher the discussion around the model, and the louder the reveal.

2. Architecture: Sparse + Linear Attention Hybrid, the First Natively Multimodal GLM-5 Model

GLM-5.3-Flash is the GLM-5 series' first natively multimodal model: 320B total parameters, 18B activated, 45 layers, pretrained on 30T tokens of multimodal corpus. Its input modalities span video, images, text, and files; the context window is 1M with 128K maximum output. It is the first open-source frontier model to combine sparse and linear attention in a hybrid architecture, and it introduces manifold-constrained hyper-connections (mHC).

The architecture-side savings, in official numbers:

DimensionGLM-5.3-FlashBaselineChange
Attention computeHybrid attentionGLM-5.33.01x lower
KV cache footprintHybrid attentionGLM-5.34.44x smaller
Activated parameters18B32B (GLM-4.5 series)down 44%
Layers4592 (GLM-4.5 series)halved

Attention compute down 3x, KV cache shrunk more than 4x, and activated parameters nearly cut in half versus the previous series - stack those three together and you have the architectural origin of the "one-fortieth the price" claim.

Then look at modalities and context. A 1M context window with 128K maximum output means you can fit an entire code repository, a full research report, or a long video in one call with room to spare; and with video, image, text, and file inputs all natively supported, you skip the multi-stage pipeline of "run a vision model first, then stitch in a text model." For agentic applications, those two together are the precondition for "one model that sees the entire work site."

3. Capability: 57 Points Level with Claude Opus 4.8, Vision Native to the Coding Loop

On Artificial Analysis's aggregate intelligence index, GLM-5.3-Flash scores 57 - level with Claude Opus 4.8, and above GLM-5.2's 53 and DeepSeek V4 Pro's official release at 53. An 18B-activated model matching flagship closed-source models is the hardest capability evidence in this launch.

Even more distinctive is Visual Coding: visual capability is natively built into the coding loop. The model can actively observe interfaces, rendering results, and interaction feedback, then keep improving - coordinating across code, browser, and graphical interfaces to complete tasks. The use cases extend to Office work, financial research, and professional documents, with deliverables including finished PPTX/PDF/DOCX/XLSX files. In other words, it is not merely "a model that can see images" - it embeds seeing into every step of the code-writing feedback loop.

4. Pricing: 1/40th of Opus, with a Cache-Hit Twist

API pricing (per million tokens): 0.8 RMB input, 2.8 RMB output, 0.23 RMB cache hit. That is 1/10 of GLM-5.3's price (as low as 1/20 during the limited-time discount window) and roughly 1/40 of Claude Opus 4.8's.

The numbers get more interesting against DeepSeek's fresh price hike. On August 17, DeepSeek raised V4-series prices and introduced peak/off-peak pricing for the first time (daily peaks at 9-12 and 14-18): V4-Flash output went from 2 RMB to 9 RMB per million tokens at peak - a 350% increase - while cached-hit input rose 400%; V4-Pro peak output hit 27 RMB (about $3.96, versus $0.87 before the hike). After the increase, V4-Flash's call volume on OpenCode fell by half (per LatePost). Here is how GLM-5.3-Flash compares to DeepSeek V4-Flash off-peak pricing:

Item (per million tokens)GLM-5.3-Flashvs. V4-Flash off-peakDuring half-price window
Input0.8 RMB53%27%
Output2.8 RMB62%31%
Cache hit0.23 RMB4.6x (V4-Flash off-peak: 0.05 RMB)--

Note the twist: the 0.23 RMB cache-hit price is actually higher than DeepSeek V4-Flash's 0.05 RMB off-peak. Coding agents routinely see cache-hit rates above 90%, meaning the bulk of input rides the cache price - for those heavy users, DeepSeek is effectively cheaper. GLM-5.3-Flash's "one-fortieth" is a flagship comparison against Opus, not a clean sweep of every rival. Before picking a side, calculate your own cache-hit rate first.

For a concrete budget reference: a team burning 100M tokens a day at a 3:1 input-output ratio pays roughly 130 RMB a day at GLM-5.3-Flash pricing (about 60 RMB for 75M input, 70 RMB for 25M output); with a 90% cache-hit rate, most input rides the 0.23 RMB tier and the bill drops further. Put the same load on Claude Opus 4.8 and the invoice multiplies by forty. That is why small and mid-sized teams watch models of this "flagship performance, lite price" class closely. For hands-on steps (Coding Plan, ZCode, and API migration paths), see the companion SOP in this batch.

5. 100,000 Domestic AI Chips: Domestic Silicon's First Large-Scale Service of Real Global Traffic

More telling than the model itself is the compute underneath. 100,000 domestic chips carried all of Ox-Alpha's online traffic during testing and the online service after launch - the first time domestic compute has served real global load at scale. Suppliers are reported to possibly include Huawei, Hygon, and Moore Threads (per LatePost/STCN citing sources; Zhipu has not commented - its official blog says "tens of thousands of domestic accelerators" plus a self-developed high-bandwidth interconnect network).

The engineering details deserve the record too: Zhipu built a dedicated inference engine on top of SGLang, using W8A8 quantization, INT8/FP8/BF16 mixed cache quantization, and Encode-Prefill-Decode (EPD) three-stage decoupled scheduling - lifting end-to-end performance 3x on the same hardware. In plain terms: W8A8 quantization compresses weights and activations to 8-bit precision, cutting memory and bandwidth overhead in step; mixed cache quantization stores KV cache entries at INT8/FP8/BF16 depending on their freshness and heat; and EPD decoupling splits encoding, prefill, and decode into separately scheduled stages so different phases no longer sit idle waiting on each other. Zhipu states that per-token hardware efficiency and cost have reached parity with mainstream NVIDIA GPUs.

The boundaries must be stated clearly: domesticating the training stage remains unsolved industry-wide, and there is no official breakdown of the 100,000-card mix. The inference-side step is real - but don't extrapolate it into "full-stack independence."

6. Open Source and the Price-War Context: MIT Weights, While the Industry Raises Prices

GLM-5.3-Flash's weights are open under the MIT license and hit Hugging Face at launch, with integrations including the ZCode coding platforms; GLM Coding Plan has fully switched to GLM-5.3-Flash with quota tripled, and 10,000 trial cards are handed out daily on a limited basis.

Put it in the context of the 2026 price war: nine days after DeepSeek's hike, Zhipu came in at one-tenth the price (as Lanjing News's headline put it) - but the broader industry is moving the other way. Zhipu itself removed the Coding Plan first-purchase discount in February and raised GLM-5-Turbo API prices 20% in March; by the end of Q1 its API pricing was cumulatively up 83%, while call volume grew 400%. Kimi K3's output price is set at 100 RMB per million tokens, more than triple the previous generation, and seven cloud and model vendors have raised prices since the start of the year (per BlockBeats tally).

In other words, the industry consensus is that "rock-bottom prices are unsustainable" - and Zhipu's play is to fight the price war with GLM-5.3-Flash, an 18B-activated model whose architecture cuts costs, while the flagship GLM-5.3 keeps its price. The low price is not generosity; it is what the new architecture buys.

There is also a practical signal buried here for buyers: when a vendor raises prices across the board while cutting one specific model to a tenth, that discount is likely not cash-burning market share chasing but architecture dividend - meaning the probability of it being clawed back soon is relatively lower. That is an inference, not a promise.

One-line closer: when 57-point intelligence is compressed to one-fortieth of Opus's price, runs on 100,000 domestic chips, and ships under MIT weights, "a cheap flagship" stops being marketing spin and becomes a quote sheet you can actually check.


References

This article is based on public reporting and official documentation (as of 2026-08-27). Prices and promotions are snapshots; the official pages prevail. Not investment advice.

This article is AI-assisted and human-edited. Last updated: 2026-08-27

Related

Frontline Hotspot

Alibaba Shows Its Qwen4 Hand Early: Qwen3.8-Flash-Next Ships 125B Weights, But No License File

On August 26 Alibaba released Qwen3.8-Flash-Next: a multimodal MoE model that doubles as an early preview of the Qwen4 architecture - the same role Qwen3-Next once played for Qwen3.5. The main model is 125B parameters with an extra 51B of N-gram embeddings, activating just 6B per token; training costs about one ninth of Qwen3.7-Plus while delivering stronger coding and office performance. Four upgrades, unpacked: GDN compresses history while QSA uses a compressed indexer to pick important context at micro-block granularity; Gated Residual widens the residual stream into four branches; the N-gram embedding table can be offloaded to host memory and overlapped with compute via async prefetch; and the optimizer switches to Muon. Native context is 262,144 tokens, extensible to 1M with YaRN. The production Qwen3.8-Flash lists at \$0.16/\$0.47 per million tokens on QwenCloud (sources differ slightly; defer to the official site). The real open question is licensing: the GitHub repo ships no LICENSE file and its license field is None, the README simply points to the Hugging Face or ModelScope model page, and the community is already asking "why isn't it Apache 2.0?" - this article marks it unconfirmed, so verify the model page before any commercial use.

Aug 30, 20266 min read
Frontline Hotspot

Zhipu Raised HK$31.4B Yet Announced It Won't Chase Profit: How Big a Future Is Tang Jie's "Touch High" Betting On?

Zhipu founder Tang Jie's July 11 internal letter "The Great Wave Is Here" launches the Touch High plan: no short-term monetization for two years, strategic investment in AGI. Just placed HK$31.4B in shares yet announced no profit-chasing-betting on four engines: long-horizon tasks, autonomous agents, full self-training, and mechanistic interpretability, straight at ASI. A breakdown of this anti-commercial-common-sense gamble and its risks.

Jul 23, 20263 min read
Frontline Hotspot

ChatGPT Images 2.5: Half the Latency, Real Consistency

OpenAI launched ChatGPT Images 2.5 on 2026-09-09: up to 50% lower latency than 2.0, better preservation of reference-photo subjects and multi-turn edit consistency; ChatGPT adds sketch mode, templates, image comments and prompt sharing; the API ships two models, Flare and Sunburst. This piece breaks down each upgrade, argues the real leap is latency plus consistency rather than raw image quality, reads the two-model split as capability tiering and pricing segmentation (analysis, not official wording), and weighs the long-term lock-in cost of closed APIs.

Sep 9, 20269 min read