Glonce

Z.ai ships GLM-5.3-Flash, nearing Claude Opus 4.8 at one-tenth the cost

modelsThe Glonce editorsHacker News🔥77Glonce rating77Impact60Novelty55Relevance95Surprise55Players50Credibility65

Z.ai introduced GLM-5.3-Flash on August 26, 2026, describing it as the first natively multimodal model in the GLM-5 series. The model has 320 billion total parameters with 18 billion active, and Z.ai says it outperforms its predecessor, GLM-5.2, across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic tasks.

On the Artificial Analysis Intelligence Index v4.1.1, GLM-5.3-Flash scores 57 at $0.045 per task (discounted), a level of intelligence Z.ai says was previously available only at roughly 10 times the cost. Against GLM-5.2, it scores 63.4 versus 46.2 on DeepSWE v1.1 and 48.8 versus 26.2 on AutomationBench v1.0.6. On Z.ai's own Code Bench v1.0, run inside Claude Code 2.1.207, it clearly beats GLM-5.2 at every reasoning effort level and, at maximum effort, nearly matches Claude Opus 4.8: 29.0 versus 29.5. On a wider six-model table spanning GLM-5.2, DeepSeek-V4-Vision-Exp, Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash, GLM-5.3-Flash leads on Terminal Bench 2.1 (84.3), Toolathlon Verified (78.4, ahead of Opus 4.8's 76.2) and GDPval-AA v2 (1773), while trailing GPT-5.6 Terra and Gemini 3.7 Flash on some vision benchmarks such as BabyVision and MVBench.

The efficiency comes from architectural changes. GLM-5.3-Flash is Z.ai's first model to combine sparse and linear attention in a hybrid architecture, aimed at cutting long-context serving costs while keeping long-context accuracy. It also adopts Manifold-Constrained Hyper-Connections (mHC) for scaling efficiency, and a technique called IndexPool that compresses four indexer key vectors into one through weighted pooling to reduce latency and memory overhead at 1 million tokens of context. Trained on a 30-trillion-token multimodal corpus, and compared with the GLM-4.5 series, GLM-5.3-Flash keeps a similar total parameter count (320B versus 355B) but nearly halves both the activated parameter count (18B versus 32B) and the number of layers (45 versus 92). Against its own sibling GLM-5.3, it cuts attention compute and KV cache size by factors of 3.0x and 4.4x respectively, though Z.ai notes its KV cache is still slightly larger than rivals Kimi-K3 and DeepSeek-V4-Flash.

Before the public release, Z.ai tested the model anonymously under the name "ox-alpha" on OpenCode and OpenRouter to gather user feedback; it says the model quickly became the most popular of the week, with all of that traffic served on Chinese AI chips. Z.ai then served GLM-5.3-Flash at cluster scale on domestically developed accelerators, building a dedicated inference engine on top of SGLang and using an Encode-Prefill-Decode disaggregated serving architecture across tens of thousands of chips. The company says its own GLM-5.3-powered infrastructure agent helped engineers write and optimize the serving kernels. Compared with its own earlier baseline on the same hardware, Z.ai reports a 3x improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost it says are comparable to mainstream Nvidia GPUs.

GLM-5.3-Flash has been rolled out to all GLM Coding Plan subscribers, giving them 3x the usable quota of GLM-5.3, available through z.ai/subscribe. Its multimodal features, including Browser Use and Computer Use for clicking through and visually verifying web pages and operating desktop apps, are available in Z.ai's ZCode tool. The model weights are public on HuggingFace, and for local deployment the model currently supports the SGLang, vLLM and TokenSpeed inference frameworks, with support for others said to be coming soon.

Key facts

  • GLM-5.3-Flash has 320 billion total parameters with 18 billion active, and Z.ai says it beats GLM-5.2 across benchmarks and real workloads at one-tenth the price.
  • On the Artificial Analysis Intelligence Index v4.1.1 it scores 57 at $0.045 per task (discounted), a level Z.ai says previously cost roughly 10 times as much.
  • On Z.ai's own Code Bench v1.0, run in Claude Code 2.1.207, GLM-5.3-Flash nearly matches Claude Opus 4.8 at maximum reasoning effort: 29.0 versus 29.5.
  • Z.ai served the model at scale on a cluster of Chinese AI chips, reporting a 3x gain in end-to-end serving performance over its own earlier baseline and per-token cost it says is comparable to mainstream Nvidia GPUs.
  • The model has rolled out to all GLM Coding Plan subscribers with 3x the usable quota of GLM-5.3, and its weights are public on HuggingFace.

Why it matters

GLM-5.3-Flash is Z.ai's first natively multimodal model and its first to combine sparse and linear attention in one architecture, a design choice aimed squarely at cutting the cost of long-context and agentic workloads rather than chasing raw capability. The headline claim, matching or nearing a frontier closed model like Claude Opus 4.8 on several coding and agentic benchmarks at roughly a tenth of the price, is the kind of cost-performance jump that reshapes which model teams default to for everyday agentic and coding work. The release also doubles as a public demonstration that a frontier-scale model can be served efficiently on domestically developed Chinese AI chips rather than Nvidia hardware.

Who it affects

Developers and companies already on Z.ai's GLM Coding Plan get the model automatically, with a tripled usable quota. Teams evaluating open-weight alternatives to Claude, GPT and Gemini for coding and agentic tasks gain another option with public weights on HuggingFace. The Chinese AI chip and inference-infrastructure ecosystem is also directly implicated, since Z.ai frames the serving results as evidence that domestic accelerators can support frontier-model inference at scale.

How to use it

GLM-5.3-Flash is already live for all GLM Coding Plan subscribers at z.ai/subscribe, with 3x the usable quota of GLM-5.3. Its multimodal Browser Use and Computer Use features, for visually verifying web pages and operating desktop applications, are available through Z.ai's ZCode tool. For self-hosting, the model weights are public on HuggingFace, and local deployment is currently supported through the SGLang, vLLM and TokenSpeed inference frameworks, with more frameworks said to be coming. Z.ai's post gives no absolute per-token API price, only the relative one-tenth comparison to GLM-5.2 and the $0.045-per-task figure on the Artificial Analysis index.

How solid is it

Every benchmark number in the announcement comes from Z.ai itself, run on Z.ai's own evaluation harnesses in most cases, including an internal framework the company used specifically to compare against DeepSeek-V4-Flash-Base for consistency. Some figures do reference third-party measurement, notably the Artificial Analysis Intelligence Index score and GDPval-AA v2, which Artificial Analysis itself evaluates. The post is detailed about evaluation settings, listing temperature, context length and harness version for each benchmark, which makes the claims checkable in principle, but none of the results have independent, third-party confirmation at the time of writing.

Risks and caveats

The strongest claim, that GLM-5.3-Flash approaches Claude Opus 4.8, rests on a narrow result: parity on one internal coding benchmark at maximum reasoning effort (29.0 versus 29.5), not a broad win. On several vision benchmarks, including BabyVision and MVBench, GLM-5.3-Flash trails both GPT-5.6 Terra and Gemini 3.7 Flash. Z.ai also discloses its own limitation: the model's KV cache is still slightly larger than rival open models Kimi-K3 and DeepSeek-V4-Flash. The company gives no absolute token pricing, no independent benchmark verification, and does not say how long the pre-release anonymous testing under the name ox-alpha ran. Two PDF documents linked in the post as examples of professional output are model-generated demonstrations, not real reports, and should not be read as factual content.

“GLM-5.3-Flash shows that frontier intelligence does not have to come at frontier cost.”

— Z.ai