OpenClaw + GLM 5.3 Flash: Opus-Class Scores at Flash Cost

What GLM 5.3 Flash Is

GLM 5.3 Flash is a mixture-of-experts model with 320 billion total parameters and about 18 billion active per token. It is the first natively multimodal model in the GLM-5 series — text, images and video go in, text comes out — and unlike GLM 5.3, which was a post-training re-tune of GLM 5.2's base, it starts from a newly trained base model.

Despite the name, it is not a distilled small sibling of GLM 5.3. It is a different architecture aimed at a different goal: keep frontier-level capability while cutting the cost of serving it. Z.ai reports that against GLM 5.3 it reduces attention compute by 3.01× and KV cache size by 4.44×.

• Hybrid sparse + linear attention — linear attention models local dependencies, a lightweight sparse indexer retrieves global context. Z.ai calls it the first open-source frontier model to combine the two.
• mHC (Manifold-Constrained Hyper-Connections) for scaling efficiency, plus IndexPool, which compresses four indexer key vectors into one to keep 1M-token contexts affordable.
• 30T-token multimodal pre-training corpus, and 45 layers against the 92 of the similarly sized GLM-4.5.
• MIT license, weights published at launch — a notable contrast with GLM 5.3, which shipped API-first with Zhipu saying weights would follow about two weeks later.

The ox-alpha Reveal

Z.ai confirms it tested the model anonymously as ox-alpha on OpenCode and OpenRouter before release to collect unbiased user feedback, and that it became the most popular model of that week. Both of those are Z.ai's own account of its pre-launch testing rather than an independent audit. Z.ai also states that all of that traffic was served on Chinese AI chips, on an SGLang-based inference stack it built for the architecture.

The Benchmark Numbers

Artificial Analysis scores GLM 5.3 Flash at an intelligence index of 57.5, with 71.5 on coding and 58.2 on agentic tasks. Placed next to Claude Opus 4.8 on the same index, the shape of the result is specific rather than a blanket win:

Read that honestly: GLM 5.3 Flash is level with Opus 4.8 on the composite index, ahead on agentic work, and behind on coding. The price gap against Opus is roughly 1/70 on input and 1/100 on output at today's promotional rate — still about 1/33 and 1/50 once it reverts to list. And it is not ahead of GLM 5.3 on any of the three: it is close, for about 5% of GLM 5.3's token cost today and around 11% at list price.

Z.ai's own reported results, which are vendor numbers and should be read as such, put it at 63.4 against GLM 5.2's 46.2 on DeepSWE v1.1, and 48.8 against 26.2 on AutomationBench. On Z.ai's internal Code Bench v1.0 at max effort it reports 29.0 against Opus 4.8's 29.5.

Pricing, and the Part That Expires

The headline price is a limited-time discount, not the list price. Z.ai's published rates per million tokens:

Budget against the list price if you are planning a workload that has to survive the promotion ending. Even at list, $0.15 / $0.50 is roughly a tenth of what GLM 5.3 costs. Cached-input storage is currently free for a limited time as well.

If you subscribe to the GLM Coding Plan instead of paying per token, Z.ai gives GLM 5.3 Flash three times the quota of GLM 5.3 under its points system, and charges half points off-peak and all weekend.

How to Run GLM 5.3 Flash on OpenClaw Launch

GLM 5.3 Flash is available through OpenRouter as z-ai/glm-5.3-flash and directly from Z.ai as glm-5.3-flash. Both routes work with a bring-your-own-key setup:

1. Deploy an agent from openclawlaunch.com — about 30 seconds. Bring-your-own-key model selection needs a paid plan; the free trial runs on a fixed set of free models, so use it to check the product, not this model.
2. In your dashboard, add your OpenRouter key (or a direct Z.AI key) under BYOK settings.
3. Open the model picker and type glm 5.3 flash. Searching reveals the live provider catalog, not just the curated shortlist, so the model appears as soon as your provider lists it — no redeploy, no config file.
4. Select it. The change applies to the running agent immediately.

Self-Hosted Configuration

Running OpenClaw on your own server? Point the default agent at the model in your openclaw.json:

{
  "models": {
    "providers": {
      "openrouter": {
        "apiKey": "sk-or-..."
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "openrouter/z-ai/glm-5.3-flash"
      }
    }
  }
}

Confirm the current slug on openrouter.ai/models before deploying rather than copying it from any guide, including this one — provider IDs shift between releases and a stale slug surfaces as a model-not-found error. Using a direct Z.AI key instead? Configure it as an OpenAI-compatible provider with its own base URL — see the custom endpoint guide.

Three Things That Will Surprise You

Thinking cannot be turned off. Reasoning is mandatory on this model — Z.ai's API accepts only thinking.type: "enabled", and OpenRouter reports reasoning as mandatory with a default effort of max. If you are used to disabling reasoning to cut latency and output tokens on cheap models, that lever does not exist here. Budget for reasoning tokens in your output cost.

Z.ai's recommended settings are unusual. Temperature 1, top_p 0.95, reasoning_effort: max, and thinking.clear_thinking: false. For streaming, enable both stream and tool_stream.

The context ceiling is provider-dependent. OpenRouter advertises up to 1,310,720 tokens across providers but the top provider serves 1,048,576 with a 131,072 max completion. Long-context agent runs should be planned against the endpoint you are actually billed by, not the maximum in the catalog.

Running the Weights Yourself

The weights are MIT-licensed and published as zai-org/GLM-5.3-Flash on Hugging Face, with official serving recipes for SGLang, vLLM, TokenSpeed and KTransformers. 320B total parameters is still a serious hosting commitment even at 18B active — the practical local path is a quantised community build rather than the full BF16 checkpoint. If you want the agent running today, an API key is the shorter road; see the vLLM guide if you intend to self-host it anyway.

When to Pick It

The interesting case is the one that was not viable before: an agent that watches a chat, calls tools all day, and reads screenshots, running continuously without the token bill becoming the reason you turn it off.

What's Next?

• OpenClaw + GLM 5.3 — the larger flagship this one undercuts
• OpenClaw + GLM 5.2
• OpenClaw + GLM — the wider GLM-4 and GLM-5 lineup
• Hermes + GLM — the same models on Hermes instead of OpenClaw
• OpenRouter setup guide
• Compare all models