Existing users must change the model name, while the $0 endpoint remains subject to unpublished usage caps.
By [Ryan Merket](/author/ryan-merket)
· Published
Primary source: [X - OrcaRouter](https://x.com/OrcaRouter/status/2096881959830229345)
Why it matters #
AI gateways can refresh the model behind a $0 endpoint quickly, but developers inherit model-ID churn, shifting latency and best-effort capacity.
OrcaRouter replaced Qwen3.8-27B with Z.ai's GLM-5.3 Flash on its free tier on September 7th, giving developers access to a larger multimodal model at $0 while requiring existing integrations to change the model name.
OrcaRouter announced the switch in a post on X, describing GLM-5.3 Flash as "more powerful" and telling users of the existing free endpoint to update their configurations. The endpoint remains rate-limited, OrcaRouter confirmed in a reply.
The change is a quality-for-speed trade on OrcaRouter's own measurements. OrcaRouter gives GLM-5.3 Flash a composite quality score of 8 out of 10, compared with 4 out of 10 for Qwen3.8-27B. GLM also carries a 1 million-token context window, versus 262,144 tokens for Qwen.
GLM was slower in OrcaRouter's seven-day measurements. OrcaRouter listed median time to first token at 7.66 seconds for GLM, compared with 1.96 seconds for Qwen. Measured generation speed was 74.2 tokens per second for GLM and 196 tokens per second for Qwen. Those figures can move with traffic and serving conditions, but they put a limit on OrcaRouter's blanket description of the replacement as an upgrade: users gain model capability and context capacity while potentially waiting longer for responses.
A newly available model becomes the free default
Z.ai published its GLM-5.3 Flash launch details on September 7th. Z.ai describes GLM-5.3 Flash as a mixture-of-experts model with 320 billion total parameters and 18 billion active parameters per token. It accepts text, images and video, and returns text.
Z.ai says the model combines sparse and linear attention to reduce the cost of processing long context. Z.ai also says it trained GLM-5.3 Flash on a 30 trillion-token multimodal corpus. Those architecture and training figures come from Z.ai, while the model weights are available under an MIT license.
For paid traffic, OrcaRouter currently lists GLM-5.3 Flash at $0.075 per million input tokens and $0.25 per million output tokens. The outgoing Qwen model is listed at $0.33 per million input tokens and $2.40 per million output tokens. Free-tier users pay neither rate, although their calls remain constrained by OrcaRouter's separate free capacity. That price gap helps explain why GLM can replace Qwen in a customer-acquisition tier. GLM offers a larger context window and higher scores in OrcaRouter's catalog while carrying a lower paid token price. OrcaRouter still has to fund the free requests, making the underlying serving cost relevant even when the customer-facing price remains zero.
Free remains best-effort
OrcaRouter's free-tier documentation says free models are ordinary catalog models exposed under separate free IDs. They retain the weights and capabilities of the paid models they shadow, but run under their own workspace-level limits.
OrcaRouter does not publish numeric ceilings for free requests per minute, requests per day or prompt size. Accounts that have never purchased credit receive a smaller daily allowance, according to the documentation. A lifetime-spending threshold raises that allowance and removes the lower tier's per-request prompt cap.
Free-tier rejections return an HTTP 429 error. A response carrying a Retry-After header indicates that a minute or daily window is full. A 429 without that header means the prompt itself exceeds the free tier's size limit and must be shortened.
Free requests also stay isolated from paid capacity. OrcaRouter says a saturated free model will not silently fall back to its paid counterpart, and free models cannot be inserted as fallback targets behind paid requests. OrcaRouter explicitly describes the tier as best-effort rather than production capacity.
The September 7th replacement therefore changes the model available inside the free allowance, not the operating terms around it. Developers get GLM-5.3 Flash's larger context window, multimodal inputs and stronger catalog scores. They also need to update a model string, account for slower measured responses and continue treating the endpoint as capacity that OrcaRouter can limit or rotate.