Discovery and access: QbitAI’s October 12 report, citing user @leo114119, placed the model in ZCode version 3.15.1, with access randomly granted to subscribers rather than automatically available to every paying user. 139 The reporter tried manually adding and switching to it unsuccessfully, and explicitly could not tell whether their Lite plan was ineligible or their account simply had not been selected for the gray release. 13

Speed claim: The roughly 490 tokens/s figure came from an early user’s observation, not an independently reproduced benchmark or guaranteed throughput. 647 If sustained, it implies about 20.4 seconds to generate 10,000 tokens—generation time alone, not total task latency. 13

Comparison: Contemporary reporting compared that observation with GLM-5.3-FlashX’s roughly 200 tokens/s peak following its September release, and with approximately 60 tokens/s attributed to Artificial Analysis’s measurement of GLM-5.3-Flash on the Z.ai API. 514 Numerically, 490 is about 2.45 times 200 and 8.2 times 60, but these are different models and measurement conditions, not a controlled head-to-head benchmark. 514

Earlier catalog entry: QbitAI reported that Pi’s model catalog had already listed GLM-5.3-Highspeed with a 1,000,000-token context window, a maximum output of 131,072 tokens, and low, high, and max reasoning-effort settings. 11 These are reported catalog specifications, not proof that every limit or setting was usable in the limited ZCode trial. 1011

Omen Alpha connection: Community speculation suggested it might be Omen Alpha, the anonymous model that appeared on OpenCode in September and had been linked to the GLM family; the proposed identity remained unverified. 56 Insufficient evidence in the supplied results confirms the specific Reddit alternatives involving an Omni or GLM-5.4-Flash variant.

Why it could matter: If the reported speed is sustained while preserving useful reasoning and coding quality, faster generation could reduce waiting across long outputs and repeated agent iterations. That is a conditional inference—not an established improvement in end-to-end task completion—and the available evidence remains a limited rollout, one reported speed observation, and catalog metadata. 6101147

Discovery and access: QbitAI’s October 12 report, citing user @leo114119, placed the model in ZCode version 3.15.1, with access randomly granted to subscribers rather than automatically available to every paying user. 139 The reporter tried manually adding and switching to it unsuccessfully, and explicitly could not tell whether their Lite plan was ineligible or their account simply had not been selected for the gray release. 13

Speed claim: The roughly 490 tokens/s figure came from an early user’s observation, not an independently reproduced benchmark or guaranteed throughput. 647 If sustained, it implies about 20.4 seconds to generate 10,000 tokens—generation time alone, not total task latency. 13

Comparison: Contemporary reporting compared that observation with GLM-5.3-FlashX’s roughly 200 tokens/s peak following its September release, and with approximately 60 tokens/s attributed to Artificial Analysis’s measurement of GLM-5.3-Flash on the Z.ai API. 514 Numerically, 490 is about 2.45 times 200 and 8.2 times 60, but these are different models and measurement conditions, not a controlled head-to-head benchmark. 514

Earlier catalog entry: QbitAI reported that Pi’s model catalog had already listed GLM-5.3-Highspeed with a 1,000,000-token context window, a maximum output of 131,072 tokens, and low, high, and max reasoning-effort settings. 11 These are reported catalog specifications, not proof that every limit or setting was usable in the limited ZCode trial. 1011

Omen Alpha connection: Community speculation suggested it might be Omen Alpha, the anonymous model that appeared on OpenCode in September and had been linked to the GLM family; the proposed identity remained unverified. 56 Insufficient evidence in the supplied results confirms the specific Reddit alternatives involving an Omni or GLM-5.4-Flash variant.

Why it could matter: If the reported speed is sustained while preserving useful reasoning and coding quality, faster generation could reduce waiting across long outputs and repeated agent iterations. That is a conditional inference—not an established improvement in end-to-end task completion—and the available evidence remains a limited rollout, one reported speed observation, and catalog metadata. 6101147