GLM-5.3-Flash explained: Benchmarks, pricing and hardware

GLM-5.3-Flash is the model that spent a week topping OpenRouter's coding charts under the name Ox Alpha. On August 26, 2026 Z.ai claimed it, published the weights on Hugging Face under an MIT license and priced the API at $0.15 per million input tokens. Since we tracked the whole stealth saga in our Ox Alpha article, this one is the reference for the model under its real name: what the benchmarks say once you separate verified numbers from vendor tables, what it costs and what it takes to run it yourself.

My short version, before the details: this is the cheapest route to near-frontier agentic coding available right now, and the benchmark claims held up under independent testing (rare enough to be worth saying). The two things to go in with open eyes about are latency and the self-hosting hardware bill.

GLM-5.3-Flash specs

From Z.ai's announcement, the model card and the API listings:

• Architecture: sparse MoE with 320B total parameters and 18B active per token, released in FP8 and BF16 checkpoints
• Context: 1M-token class window with 131,072 max output on the API listings (the stealth listing showed 1,048,576; the model card's own evaluations cite 300K)
• Modalities: natively multimodal, the first in the GLM-5 series; text and image input on the model card, with video input listed on the API side
• Reasoning: always on, with a reasoning_effort parameter (low, high, max) defaulting to max; it can't be disabled
• Capabilities: tool calling and structured JSON output; officially supported languages are English and Chinese

Z.ai also describes a hybrid attention design (linear plus sparse) that it credits with roughly 3x less attention compute and a 4.4x smaller KV cache at long context versus GLM-5.3. Those are vendor architecture claims without independent measurement yet, so treat them as engineering direction rather than gospel. The knowledge cutoff is undisclosed anywhere public, which fits a lab that spent a week hiding the model's name.

GLM-5.3-Flash benchmarks

The official numbers

Z.ai's launch table puts GLM-5.3-Flash at 63.4 Pass@1 on DeepSWE v1.1 (GLM-5.2 scored 46.2), 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0 and GPT-5.6 Terra's 87.4, 55.3 on Humanity's Last Exam with tools and 48.8 on AutomationBench, up from 26.2 for GLM-5.2. On Z.ai's in-house Code Bench it reaches 29.0 at max effort versus 29.5 for Opus 4.8, which is the source of every "matches Claude at a tenth of the price" headline you've seen this week.

What independent testing confirms

This is the part that surprised me. The official DeepSWE leaderboard now carries a glm-5.3-flash entry at 63% with a 4-point margin, matching Z.ai's self-reported 63.4. The community run against the anonymous Ox Alpha endpoint before the reveal landed at 58.4% on the same suite, inside the same confidence band. And Artificial Analysis independently scores it 57 on their Intelligence Index, frontier territory for a model at this price point. Vendor benchmark tables usually shrink on contact with independent harnesses. This one didn't, and after the pre-reveal week where a fake 80% figure went viral, the real numbers holding steady counts for something.

Against DeepSeek and Kimi

The comparisons circulating pit GLM-5.3-Flash against DeepSeek-V4-Vision-Exp (63.4 vs 59.3 on DeepSWE, 48.8 vs 38.8 on AutomationBench, and a clean sweep of the vision benchmarks) and against Kimi K3, where the picture splits: K3 leads on deep reasoning suites like Terminal-Bench (88.3 vs 84.3) while GLM-5.3-Flash edges ahead on tool-heavy automation (Toolathlon 78.4 vs 76.5). One honest caveat: these tables mix each vendor's self-reported numbers, so read them as positioning rather than a neutral lab's verdict. The practical summary I'd give a team choosing today: pick K3 for long-horizon reasoning, GLM-5.3-Flash for agents that live inside tools, and check the price column before either.

GLM-5.3-Flash API pricing

Z.ai's list pricing is $0.15 per million input tokens, $0.50 per million output, $0.03 for cached input and free cache storage during launch. Until September 9, 2026 a promotion halves the token rates ($0.075 in, $0.25 out, $0.015 cached), and OpenRouter's z-ai/glm-5.3-flash route currently mirrors the promo prices exactly, served by a dozen providers.

Two cost mechanics worth knowing. Reasoning always runs and those thinking tokens land on your output bill per the API guides, so reasoning_effort is your main cost lever: max for hard agentic tasks, low for classification and extraction where deep thinking just burns money. And cached input at a fifth of the fresh-input price rewards agent setups with long stable system prompts, which describes most serious coding harnesses. The free stealth endpoint is gone, by the way; stealth/ox-alpha now returns a 404 with a thank-you note.

Hardware requirements for self-hosting

Here's the section where I get to correct the most repeated mistake of the week. "18B active parameters" does not mean the model fits in 18B-class hardware. Active parameters cut per-token compute, so you get better speed and cheaper serving. The full 320B checkpoint still has to sit in memory: every expert must be loaded because the router picks between them at runtime, per token.

The real numbers, from Z.ai's own serving recipe and YottaLabs' hardware analysis: the FP8 checkpoint is about 306 GiB of weights with roughly 386 GiB of VRAM recommended for the default deployment, and BF16 doubles that to about 772 GiB. Two H200s (282 GiB combined) don't clear the bar, so there's no two-GPU entry point; the practical floor is an 8-GPU Hopper-class node. For scale, the biggest card in our GPU VPS lineup carries 96 GB of VRAM, and even that holds under a third of the FP8 weights. This is datacenter self-hosting, full stop.

Quantized versions

Quantization moves the bar but doesn't remove it. A 4-bit GGUF is around 160 GB before overhead, which brings a 2-4 GPU setup into range with quality trade-offs you should measure rather than assume. Unsloth publishes 1-bit and 3-bit GGUF builds it says run on machines with 100 to 128 GB of combined RAM and VRAM, Mac Studio and DGX-class territory; those are community builds at aggressive compression, fine for experiments and a stretch for production. Engine support is unusually good for a day-one release: vLLM, SGLang, KTransformers and TokenSpeed are covered in the official docs, with llama.cpp support landing through the community GGUF builds.

My honest recommendation for almost everyone: use the API. At promo pricing, a million output tokens costs a quarter dollar; the 8-GPU node that self-hosting needs costs more per day than most teams will spend on the API per month. Self-hosting makes sense in two cases, sustained high volume or data that can't leave your infrastructure, and the second one matters here since the API is served from Z.ai's stack in China.

Is GLM-5.3-Flash fast?

The name oversells it. "Flash" describes the cost and compute profile, not the feeling of using it. Artificial Analysis measured a median time-to-first-token around 42 seconds on their reasoning-heavy suite, because mandatory reasoning at the max default thinks before it speaks. Other harnesses with lighter prompts report first tokens in a second or two and throughput anywhere from 40 to over 130 tokens per second depending on provider and load. The stealth week's chronic slowness has eased as traffic spread across providers, but the pattern stands: for overnight agents nobody notices, for interactive loops you'll want reasoning_effort turned down. Also worth budgeting for: during the preview OpenRouter logged a tool-call error rate around 4.5%, and the pre-reveal benchmark run lost 11 of its 47 failures to responses that skipped the tool call entirely, so build retries into any agent pipeline.

The Ox Alpha backstory and the chips

The stealth week is a story on its own (six days, 62 trillion tokens, a debunked viral benchmark and a tokenizer that gave the game away) and the full timeline lives in the Ox Alpha article linked above. The reveal's headline beyond the model itself: Z.ai says the entire preview ran on a cluster of 100,000 domestically produced Chinese chips at cost-per-token it calls comparable to mainstream Nvidia GPUs, and its shares jumped on the news. Which chips exactly, nobody outside Z.ai knows; reporting points at Huawei, Moore Threads and Hygon as candidates, unconfirmed. Whichever way you read the geopolitics, an MIT-licensed frontier-class model served at these prices resets what the open-weight tier is expected to deliver.

Agent builders got the best deal out of all of this. Any harness that speaks an OpenAI-compatible API can point at glm-5.3-flash today, and if you're running something like Hermes on a small server, our Hermes agent guide pairs naturally with a model this cheap per token. The stealth week proved the demand; the pricing makes it sustainable. What it does to DeepSeek's next release is the sequel worth watching.