GLM models on the same tests, re-tested on a fixed schedule

Places are the leaderboard's own: walking down a board, a model joins the group above it unless that group's leader is ahead by more than the measurement can resolve, so “joint” places are normal. The same boards hold every vendor's models; sign in for the full leaderboard.

Prices are list prices per 1M tokens on api.z.ai, cache-miss rate, as the leaderboard's price view quotes them. What a correct result costs over the last 30 days is on the best-value AI model; response times are on the fastest AI model.

On the Coding board, re-tested every four hours, the best-placed GLM model in the last 24 hours is GLM-5.3: joint 16th of the 24 models on the Coding board in the last 24 hours, scoring 90 ± 0.2.

The coding suite gives a model a small working project and a bug report (7 tasks a sweep) and grades the fix by running the project's own tests, including hidden tests the model never sees. Every vendor's models get the same tasks: see the best LLM for coding across all vendors and how coding is scored.

On the Reasoning board, re-tested daily, the best-placed GLM model in the last 24 hours is GLM-5.3: joint 15th of the 22 models on the Reasoning board in the last 24 hours, scoring 91 ± 2.5.

The reasoning suite runs 4 long multi-turn working sessions a day, scored on correctness, recovery after a failed step and continuity: whether the model still honours rules stated earlier in the session, checked by running code rather than by a judge model. See the best LLM for reasoning across all vendors and how reasoning is scored.

On the Tool use board, re-tested daily, the best-placed GLM model in the last 24 hours is GLM-5.3: joint 1st of the 22 models on the Tool use board in the last 24 hours, scoring 91 ± 2.5.

The tool-use suite runs 9 tasks in sandboxed machines, each with tools to call: either the job is done at the end or it is not. It is scored on task completion, tool selection, parameter accuracy, efficiency, error handling, context awareness and safety. See the best AI model for agents across all vendors and how tool use is scored.

The 7-day history, every episode and the other five providers are on Z.ai API status and the provider status hub. When a provider stops answering, the Smart Router's fallback order moves traffic to the next model on the board.

The GLM chat app is not the Z.ai API: we probe the API, which is what applications and the Smart Router call. A provider can change a model behind the same API name, and without continuous re-testing that change is invisible to the people relying on it. AI Stupid Level compares each GLM model with its own past: every suite has its own Page-Hinkley (a CUSUM-family test) change-point detector, which separates a sustained decline from ordinary run-to-run noise, and the hourly canary probe watches for a sudden collapse. On the Combined board over the last 24 hours, of the 1 GLM model with a current score, 1 is holding steady. Searches for “is GLM getting worse” and “AI model nerfed” end at the same place: the current answer, model by model, is on Is GLM getting worse?, and the method is on AI drift detection.

GLM-5.3, the best-placed GLM model on the Combined board, against each other vendor's best-placed model:

• GLM-5.3 vs GPT-5.6 Luna
• Claude Fable 5.1 vs GLM-5.3
• Gemini 3.8 Flash vs GLM-5.3
• DeepSeek V4 Pro vs GLM-5.3
• GLM-5.3 vs Kimi K3.1 Pro

Other vendors on the same tests: OpenAI models, Anthropic models, Google models, DeepSeek models and Moonshot AI (Kimi) models. The compare hub lists every family, and the best AI model ranks them across vendors.

Every model on the board gets the same tasks on the same schedule: coding every four hours, reasoning and tool use daily, with an hourly canary probe. GLM models are called through api.z.ai on our own account, exactly as an application would call them, so a score describes the model as it is served, not as it was announced.

The coding suite (7 tasks) grades a bug fix by running the project's own tests, including hidden tests the model never sees. The reasoning suite (4 sessions) scores long multi-turn working sessions on correctness, recovery and continuity. The tool-use suite (9 tasks) runs in sandboxed machines, where the job is either done or not. The Combined score weights coding 50%, reasoning 25% and tool use 25%.

The methodology is public — every weight, statistical test and drift constant. The test tasks are private, so no lab can train on them. Read the methodology and how drift detection works.