Leaked KingBench Scores Show Zhipu GLM Models Beating Opus 5.5
Leaked KingBench 4 leaderboard images suggest Zhipu AI's upcoming GLM models—including Pro, Med, and Small variants—achieve scores of 59.4%, 57.1%, and 54.0% respectively, surpassing Anthropic's Opus 5.5 and rivaling Claude Fable 5.2, with community pixel analysis hinting at model names like GLM-Nano and GLM-Next.
Conflicting Leaked Scores for Suspected Zhipu GLM Model
Two circulating KingBench 4 leaderboard screenshots show different scores for what appears to be the same unreleased Zhipu model. The first image, reconstructed by a community member from a redacted leaderboard, shows a score of 59.4% , placing it 1.6 percentage points above Opus 5.5 (57.8%) but below Claude Fable 5.2 (70.0%). A second image, posted in the comments, shows a dramatically higher 70.1% , just 0.1 points above Claude Fable 5.2 .
Three Redacted Entries with Pro, Med, Small Suffixes
The original unredacted KingBench 4 leaderboard, shared by maintainer AICodeKing in a reply to developer Theo (t3.gg), contains 16 entries . Three consecutive rows are redacted; only the suffixes Pro, Med, and Small are visible. The benchmark consists of 6 programming problems worth 60 points total ; percentages reflect the fraction of total points earned.
Community Pixel Analysis Suggests GLM Prefix
To recover the hidden prefix, a community member rendered GLM-Name in a similar font and layout, applied Gaussian blur, and performed a pixel-by-pixel comparison with the redacted image. The pixel residue was reported to be less than 1/255 . Independent AI-assisted OCR also suggested the first three letters are GLM . Subsequent cleaned-up versions of the leaderboard circulating online show the names GLM-Nano , GLM-Pro , GLM-Med , and GLM-Small , while commenters also speculated GLM-Next .
Detailed Scores for Pro, Med, Small and GLM-5.3 Baseline
GLM-Pro: 59.4% (2nd place), behind Claude Fable 5.2 (70.0%), ahead of Opus 5.5 (57.8%).
GLM-Med: 57.1%, just behind Opus 5.5, ahead of GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon.
GLM-Small: 54.0% (9th place), between Haiku 5.5 and Kimi K3.1 Pro; ahead of Gemini 4 Carbon, Muse Wonder Full 1.0, DeepSeek V4.1 Flash.
GLM-5.3 (existing model): 45.8%, 8.2 points below GLM-Small .
The gap between GLM-5.3 and GLM-Small suggests a substantial generational improvement if Small represents the next-generation small variant.
Developer Signal and Historical Release Pattern
OpenCode developer dax commented that after DeepSeek's scaling, "next is GLM" and replied "full" when asked whether a flash or full version would come first. While dax did not mention a version number, the community linked this to the leaked trio of models.
Historical analysis of AICodeKing's posts since the GLM-4 era shows five instances where he shared internal leaderboards before official announcements. The average lead time from his post to official release is 3.8 days . For GLM-5.2, he posted the KingBench 3 internal leaderboard roughly three days before the official announcement .
Based on this pattern, commenters speculated a release within the week, with some betting on November 19 . As of the article date, Zhipu has not announced a release date.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
