@zai/glm-5.3-flash is live on Brainbase

GLM-5.3-Flash is now available on Brainbase as @zai/glm-5.3-flash. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

Z.ai released GLM-5.3-Flash on August 26 as the first natively multimodal model in the GLM-5 series: a 320B-total, 18B-active Mixture-of-Experts model with a 1,048,576-token context window, text, image, and video input, and open weights on Hugging Face under the MIT License.

#From Ox Alpha to GLM-5.3-Flash

Most model launches start with a press release. This one started with a mystery. On August 20, an anonymous model called stealth/ox-alphaappeared on OpenRouter and OpenCode with a million-token context window and a price of zero, and within days it was the most-called model on those platforms. Community fingerprinting pointed at Z.ai's GLM line, but nobody claimed it.

The August 26 announcement confirmed the theory — and revealed the preview's scale: 503K users, 13.12M completed sessions, and 44 trillion tokens served, all from inference running entirely on Chinese AI chips.

#Why GLM-5.3-Flash matters for agents

The interesting part is not the parameter count — it is what the architecture buys an agent in production. With 18B active parameters, 3× less attention compute, and a 4.4× smaller KV cache than GLM-5.3, the model keeps long agent sessions cheap to serve even as context fills up. A million tokens of context means a repository, its documentation, and a long trajectory of tool calls fit without aggressive compaction.

Native multimodality changes what those agents can verify. Z.ai trained the model on a 30-trillion-token multimodal corpus and taught it to inspect rendered interfaces, gameplay, and 3D output, then assess and revise its own work from visual feedback. The same loop covers documents, spreadsheets, presentations, and dashboards — the artifacts knowledge-work agents actually produce.

#GLM-5.3-Flash benchmarks

On Z.ai's published results, GLM-5.3-Flash beats GLM-5.2 across the board at roughly one-tenth the price: DeepSWE v1.1 rises from 46.2 to 63.4 and AutomationBench nearly doubles from 26.2 to 48.8. On Terminal-Bench 2.1 it reaches 84.3, within a point of Claude Opus 4.8, and on Z.ai's internal Z.ai Code Bench v1.0 — run inside Claude Code — it lands within half a point of Opus 4.8. These figures are vendor-reported; select a metric below to compare them.

Independent testing from Artificial Analysis scores GLM-5.3-Flash at 57 on its Intelligence Index — well above the median of 27 for open-weights models of similar size — at a measured $0.09 per Intelligence Index task and an output speed of 48.7 tokens per second on Z.ai's own API.

#GLM-5.3-Flash pricing and context

Pricing is where GLM-5.3-Flash separates from the frontier pack. Standard API rates are an order of magnitude below comparable proprietary models, and a launch promotion halves them through September 9. Because the weights are MIT-licensed, teams can also serve the model themselves with SGLang or vLLM when unit economics or data boundaries demand it.

These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. At $0.50 per million output tokens, even reasoning-heavy trajectories that would be routing decisions on a frontier model become rounding errors here.

#Run GLM-5.3-Flash on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs GLM-5.3-Flash inside the OpenCode harness — the same harness the Ox Alpha preview was most popular on — in a managed sandbox:

curl https://api.brainbaselabs.com/v2/threads \
  -H "Authorization: Bearer $BRAINBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": {
      "harness": "opencode",
      "model": "@zai/glm-5.3-flash"
    },
    "input": "Build the dashboard from the attached mockup, run it, and iterate until the rendered page matches."
  }'

#Where GLM-5.3-Flash fits

GLM-5.3-Flash is the obvious candidate for high-volume agent workloads: triage, code review at scale, document and spreadsheet processing, and any pipeline where a frontier model was chosen by default rather than by need. Its multimodal self-verification also makes it a strong pick for UI-building agents that need to look at what they shipped.

Forty-four trillion tokens of anonymous traffic in one week suggests developers had already decided this model was useful before they knew what it was. Now that it has a name, run it against your production traces. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model is the only variable.