GLM 5.3 Flash API, OpenAI-compatible
Z.ai’s GLM 5.3 Flash — the first natively multimodal model in the GLM-5 series, a 320B-parameter MoE activating just 18B per token — with a 1,048,576-token context window and image input we verified with our own probe before selling it. Served on Cloudflare Workers AI infrastructure through one OpenAI-compatible endpoint: prepaid USD credits, per-token billing you can audit, and 40% off the list price for the launch period.
What you get
- Context window
- 1,048,576 tokens (input + max_tokens)
- Architecture (per Z.ai)
- MoE, 320B total / 18B active parameters; hybrid linear + sparse attention built to cut long-context serving cost — which is exactly what the pricing below reflects
- Image input
- Yes — verified 2/2 by our own probe on 2026-08-27, base64 data URIs
- Function calling
- Yes — verified in our contract tests, tools called with correct schema
- Reasoning
- Yes — streamed as reasoning_content, billed once at output rate; effort low/medium/high/xhigh all accepted
- Cached input
- A real published rate — see the caching note below
- Upstream ID (public)
- @cf/zai-org/glm-5.3-flash
- Privacy
- Runs on Cloudflare infrastructure; we log metadata only — never prompt or response content
Pricing (40% off list, applied automatically)
- Input
$0.15$0.09- Output
$0.50$0.30
@cf/zai-org/glm-5.3-flash
Struck price = list. The cached-input rate is a real published rate and we saw real hits billed at it in our measurements — but hits are not yet consistent upstream, so treat the discount as a bonus, not a budget line. Live prices for all eight models are on the models page, synced from the billing database that rates your requests.
Measured, not quoted
Our own measurements, 2026-08-27 — the day we added the model. Live status and per-model numbers stay on the models page.
~1.6 s
time to first token (p50, 5 streams, none hung)
~80 tok/s
streaming throughput (p50)
31–47 ms / 1K tokens
prefill rate measured at 150K–600K input
~19 s
first token on a real 600K-token prompt — use streaming for large jobs
Ship in two lines of config
If your code already talks to the OpenAI API, you only change base_url and the key. The model id is glm-5.3-flash; the upstream id is public.
from openai import OpenAI
client = OpenAI(
base_url="https://api.clfaigateway.dev/v1",
api_key="sk-gw-...", # app.clfaigateway.dev
)
resp = client.chat.completions.create(
model="glm-5.3-flash",
messages=[{"role": "user", "content": "Summarize this repository..."}],
stream=True,
)
for chunk in resp:
print(chunk.choices[0].delta.content or "", end="")GLM 5.3 Flash on the gateway — common questions
The window counts input plus your max_tokens budget. At this size prefill is real work — we measured roughly 31–47 ms per 1,000 input tokens, so a 600K-token prompt takes about 19 seconds before the first token. Always stream large requests, and expect retryable capacity errors at peak times on very large prompts.
Context & caching docs →Yes — and we checked before selling it. Same probe as the rest of the catalog: two solid-colour images, the model must identify both. GLM 5.3 Flash passed 2/2 on 2026-08-27. Images go in as base64 data URIs in the standard content array; remote http(s) links are refused up front. Z.ai additionally reports strong document-and-chart understanding (OfficeQA Pro 62.4 on scan-only PDFs, CharXiv Reasoning 89.4) — we have not probed OCR quality ourselves yet, so test that on your own material first.
Which models read images →Z.ai’s own published benchmarks put it ahead of GLM-5.2 across the board — 63.4 vs 46.2 on DeepSWE v1.1, 48.8 vs 26.2 on AutomationBench, 78.4 vs 59.9 on Toolathlon Verified — while approaching much larger frontier models on coding and agentic work. Those are the vendor’s numbers, not ours; what we verified ourselves is the table above. One thing that is ours: on our price list, its input rate is about 9× cheaper than glm-5.2’s.
Compare live prices →The rate is real and published upstream, and our measurements saw genuine cache hits billed exactly at it. But identical repeats also missed the cache in the same run — hits are not yet consistent, so we say so instead of promising savings. When a hit lands you pay the cached rate automatically.
How caching behaves →We added it the day it appeared, measured everything we publish, and saw occasional transient upstream errors during testing — the kind that succeed on retry. The gateway translates those into a clean retryable error and fails over where it can. If reliability shifts, the status page and this page will say so.
Prepaid credits, billed per actual token from upstream usage — never estimates. Reasoning tokens are part of output and billed once. Every request shows its exact cost in your usage log, and the ledger adds up to the nano-dollar.
Billing docs →