Models
Every AI model we benchmark on the same tests, on a fixed schedule
AI Stupid Level gives every model on the board the same coding, reasoning and tool-use tasks: coding every four hours, reasoning and tool use daily, with an hourly canary probe. The methodology is public — every weight, statistical test and drift constant. The test tasks are private, so no lab can train on them.
Scores are ASL's own, from the 24-hour boards; places carry the leaderboard's statistical ties.
- 24
- models on the 24-hour boards, from 6 vendors
- 22
- ranked on the Combined board, with statistical ties
- 6 Feb 2026
- tracked since, for the longest-tracked model on the roster
01
Which AI models are on the board?
24 models from 6 vendors are on the 24-hour boards; 22 are ranked on the Combined board. 15 models share first place; the board lists DeepSeek V4.1 Flash first, joint 1st of the 22 models on the Combined board in the last 24 hours, scoring 95 ± 2.2.
As of 26 Oct 2026, 09:32 UTC · 24-hour boards · score, then place in brackets; “=” marks a statistical tie
The tables below list every model by vendor, in Combined board order: the Combined score (coding 50%, reasoning 25%, tool use 25%) with its standard error and place, then each suite's score and place. A model whose reasoning or tool-use runs are funded by the community is shown but not ranked on the Combined board. Each name opens the model's own page; each vendor heading has a page of its own.
Free account: every ranking over 24 hours, 7 days and 30 days, a 3-model watchlist with email alerts, 1,000 routed requests a month and a Data API key. No card.
Already have an account? Log in
02
Which GPT models are benchmarked?
8 OpenAI (GPT) models are on the roster, called through api.openai.com on our own account, as an application would call them.
| Place | Model | Score | Trend | Coding | Reasoning | Tool use |
|---|---|---|---|---|---|---|
| =1 | GPT-5.6 Luna | 94 ± 2.2 | holding steady | 96 (=1) | 99 (=1) | 86 (=1) |
| =1 | GPT-6 Sol | 94 ± 2.3 | holding steady | 95 (=1) | 98 (=1) | 89 (=1) |
| =1 | GPT-5.6 Sol | 93 ± 2.2 | holding steady | 93 (=6) | 100 (=1) | 85 (=1) |
| =1 | GPT-6 Astra | 93 ± 2.3 | holding steady | 93 (=6) | 99 (=1) | 86 (=1) |
| =1 | GPT-6 Luna | 92 ± 2.3 | holding steady | 92 (=6) | 97 (=1) | 87 (=1) |
| =1 | GPT-5.6 Terra | 90 ± 2.3 | holding steady | 92 (=6) | 98 (=1) | 78 (=14) |
| =16 | GPT-5.5 | 82 ± 2.3 | holding steady | 94 (=1) | 60 (=17) | 78 (=14) |
| =16 | GPT-6.1 Sol | 81 ± 2.3 | holding steady | 93 (=6) | 58 (=17) | 81 (=14) |
Every OpenAI model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; OpenAI API status, checked every 10 minutes; and is ChatGPT getting worse.
03
Which Claude models are benchmarked?
9 Anthropic (Claude) models are on the roster, called through api.anthropic.com on our own account, as an application would call them.
| Place | Model | Score | Trend | Coding | Reasoning | Tool use |
|---|---|---|---|---|---|---|
| =1 | Claude Opus 5.5 | 94 ± 2.3 | holding steady | 95 (=1) | 99 (=1) | 86 (=1) |
| =1 | Claude Sonnet 5.5 | 92 ± 2.2 | holding steady | 93 (=6) | 97 (=1) | 86 (=1) |
| =1 | Claude Fable 5.1 | 92 ± 2.2 | holding steady | 88 (=18) | 100 (=1) | 90 (=1) |
| =1 | Claude Opus 5 | 90 ± 2.3 | holding steady | 87 (=18) | 98 (=1) | 89 (=1) |
| =1 | Claude Sonnet 5 | 89 ± 2.2 | holding steady | 89 (=15) | 90 (=15) | 88 (=1) |
| =19 | Claude Opus 4.8 | 78 ± 2.3 | holding steady | 88 (=18) | 59 (=17) | 78 (=14) |
| =19 | Claude Opus 4.7 | 78 ± 2.2 | holding steady | 93 (=6) | 53 (=20) | 72 (=20) |
| =19 | Claude Opus 4.6 | 76 ± 2.2 | holding steady | 90 (=15) | 52 (=20) | 70 (=20) |
| =19 | Claude Sonnet 4.6 | 74 ± 2.3 | holding steady | 90 (=6) | 46 (22) | 70 (=20) |
Every Anthropic model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; Anthropic API status, checked every 10 minutes; and is Claude getting worse.
04
Which Gemini models are benchmarked?
3 Google (Gemini) models are on the roster, called through generativelanguage.googleapis.com on our own account, as an application would call them.
| Place | Model | Score | Trend | Coding | Reasoning | Tool use |
|---|---|---|---|---|---|---|
| =1 | Gemini 3.8 Flash | 90 ± 2.3 | holding steady | 92 (=6) | 98 (=1) | 77 (=14) |
| =16 | Gemini 3.5 Flash Lite | 86 ± 2.2 | holding steady | 85 (22) | 96 (=1) | 78 (=14) |
| not ranked | Gemini 4 Argon | holding steady |
Every Google model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; Google API status, checked every 10 minutes; and is Gemini getting worse.
05
Which DeepSeek models are benchmarked?
2 DeepSeek models are on the roster, called through api.deepseek.com on our own account, as an application would call them.
| Place | Model | Score | Trend | Coding | Reasoning | Tool use |
|---|---|---|---|---|---|---|
| =1 | DeepSeek V4.1 Flash | 95 ± 2.2 | holding steady | 96 (=1) | 99 (=1) | 90 (=1) |
| =1 | DeepSeek V4 Pro | 90 ± 2.2 | holding steady | 87 (=18) | 95 (=1) | 91 (=1) |
Every DeepSeek model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; DeepSeek API status, checked every 10 minutes; and is DeepSeek getting worse.
06
Which Kimi models are benchmarked?
1 Moonshot AI (Kimi) model is on the roster, called through api.moonshot.ai on our own account, as an application would call it.
| Place | Model | Score | Trend |
|---|---|---|---|
| not ranked | Kimi K3.1 Pro | holding steady |
Every Moonshot AI model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; Moonshot AI API status, checked every 10 minutes.
07
Which GLM models are benchmarked?
1 Z.ai (GLM) model is on the roster, called through api.z.ai on our own account, as an application would call it.
| Place | Model | Score | Trend | Coding | Reasoning | Tool use |
|---|---|---|---|---|---|---|
| =1 | GLM-5.3 | 91 ± 2.2 | holding steady | 90 (=15) | 91 (=15) | 91 (=1) |
Every Z.ai model on the same tests, with its place on Coding, Reasoning and Tool use and the family's drift state; Z.ai API status, checked every 10 minutes.
08
How do I compare two models?
Every pair has a head-to-head page: both models on the four boards with the leaderboard's statistical ties, list price per 1M tokens, cost per correct result over the last 30 days and drift state.
The compare hub lists every family and the curated pairs; each model page links its closest rivals and each vendor's best. For a decision by job, the best AI model ranks the leaders of every board, with pages for coding, reasoning, agents and tool calling, value and speed. Whether a model is slipping is answered on is AI getting worse.
One OpenAI-compatible endpoint, your own provider keys, 1,000 requests a month free · no card
09
How are the models tested and ranked?
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. The coding suite grades a bug fix by running the project's own tests, including hidden tests the model never sees; the reasoning suite scores long multi-turn working sessions on correctness, recovery and continuity; the tool-use suite runs in sandboxed machines, where the job is either done or not. The Combined score weights coding 50%, reasoning 25% and tool use 25%; a suite that is missing or out of date drops out and the rest are reweighted.
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 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.
10
Questions people ask about the models we benchmark
How many AI models does AI Stupid Level benchmark?
24 models from 6 vendors are on the roster: OpenAI (GPT) (8), Anthropic (Claude) (9), Google (Gemini) (3), DeepSeek (2), Moonshot AI (Kimi) (1) and Z.ai (GLM) (1). 24 of them have a score on the 24-hour boards. Each has a page of its own with its place on every board, its drift state and its head-to-head comparisons. The list grows as models are released and shrinks when a provider retires one.
Which AI model has the highest Combined score?
DeepSeek V4.1 Flash: joint 1st of the 22 models on the Combined board in the last 24 hours, scoring 95 ± 2.2. 15 models share first place: their scores are closer together than the measurement can separate. The Combined score weights coding 50%, reasoning 25% and tool use 25%. The leaders of every board are on the best AI model.
How often is each model re-tested?
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. The methodology is public — every weight, statistical test and drift constant. The test tasks are private, so no lab can train on them. Every weight is on the methodology page.
How are the places decided?
Every score carries a standard error from its own run-to-run repeatability. Walking down a board, a model joins the group above it unless that group’s leader is ahead by more than 1.96 times their combined standard error, so places such as 1, 1, 1, 4 are normal and a shared place is written "joint" (an "=" in the tables). A model whose reasoning or tool-use runs are funded by the community is shown but not ranked on the Combined board. The rule is on the methodology page.
See every board, or let the router choose
Free account: every ranking over 24 hours, 7 days and 30 days, a 3-model watchlist with email alerts, 1,000 routed requests a month and a Data API key. No card.
Free account · 24 hours, 7 days and 30 days · no card
Already have an account? Log in
Routes across every vendor on the board, on your own provider keys · 1,000 requests a month free