April 3rd news, yesterdayZhipu GLM-51 is on line, no press meetings, no technical reports, and is available directly to all users of the GLM Convention Plan (Lite, Pro, Max)。
GLM-51 SCORED 45.3 IN THE OFFICIALLY PUBLISHED PROGRAMMING CAPACITY ASSESSMENTand Claude Opus 4.6 for 47.9 points only 2.6 minutesI DON'T KNOW. COMPARED TO THE PREVIOUS GLM-5 35.4 POINTS, THE SINGLE-STAGE INCREASE WAS NEARLY 10 POINTS AND THE INCREASE WAS APPROXIMATELY 281 TP3T。
GLM-5.1 CORE COMPETENCY ENHANCEMENT FOCUSES ON:
Programming scores 45.3, SWE-bench Verified scores 77.8%, the gap with Opus 4.6 is within 3 percentage points
Supports 200K tokens context window and 131K tokens maximum output length
High-profile performance in Agent scenarios such as BrowneComp, MCP-Atlas complex tools, 2-Bench multi-tool integrated implementation
THE HALLUCINOGENIC RATE DROPPED FROM 38.61 TP3T IN THE PREVIOUS GENERATION TO 34.11 TP3T。
In addition, yesterday, a multi-mode Coding Base Model GLM-5V-Turbo was announced:
original supports multi-modular input such as pictures, videos, designs, etc., context window 200k
Core competencies: to understand the script/intercept directly to generate operational codes (front end recapturing)
Fits to Claude Code, AutoClaw depth and supports complete GUI Aget closed rings
High performance on GUI benchmarks such as AndroidWorld, WebVoyager and so forth, with the pure text programming capability being synchronized。