BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases9

IFBench

Precise instruction following on unseen constraints. Run by Artificial Analysis.

As of 25 Oct 2026, Grok 4.3 leads IFBench on BenchLeader with 83.3%, ahead of MiniMax M3 at 82.9%, across 423 model configurations with a published result.

Published by
Artificial Analysis
Category
Instruction following
Index weight
1.0
Models
423
Data as of
25 Oct 2026
Newest result
23 Oct 2026

Source: Artificial Analysis (artificialanalysis.ai). Data taken from the public leaderboard.

What the test looks like

IFBench tests precise instruction following with constraints the model has not seen before, such as unusual formatting or counting rules, applied to ordinary writing tasks.

How it is scored

Percent of instructions satisfied, verified programmatically; run by Artificial Analysis.

What to keep in mind

It measures compliance with explicit rules, which is what makes a model reliable in pipelines, but not judgment about what the user meant.

Top 15
  1. 1Grok 4.3 (medium)83.3%
  2. 2MiniMax M382.9%
  3. 3Nemotron 3 Ultra 550B A55B (thinking)81.4%
  4. 4Grok 4.3 (high)81.3%
  5. 5Grok 4.20 (thinking)81.2%
  6. 6Grok 4.3 (low)81.0%
  7. 7Qwen3.7 (max)80.5%
  8. 8Nemotron Cascade 2 30B A3B80.4%
  9. 9MiMo-V2.5-Pro (thinking)79.9%
  10. 10Nova 2.0 Pro Preview (low)79.6%
  11. 11DeepSeek V4 Flash 0420 (max)79.2%
  12. 12Nova 2.0 Pro Preview (medium)79.0%
  13. 13Qwen3.5 397B A17B (thinking)78.8%
  14. 14Gemini 3 Flash (thinking)78.0%
  15. 15Qwen3.7 Plus78.0%
423 of 423
#
1Grok 4.3medium83.3%58.72026-04-30
2MiniMax M382.9%56.42026-06-01
3Nemotron 3 Ultra 550B A55Bthinking81.4%55.72026-06-04
4Grok 4.3high81.3%57.02026-04-30
5Grok 4.20thinking81.2%58.62026-04-07
6Grok 4.3low81.0%56.92026-04-30
7Qwen3.7max80.5%60.62026-05-19
8Nemotron Cascade 2 30B A3B80.4%49.62026-03-19
9MiMo-V2.5-Prothinking79.9%57.02026-04-22
10Nova 2.0 Pro Previewlow79.6%49.32025-11-27
11DeepSeek V4 Flash 0420max79.2%57.02026-04-24
12Nova 2.0 Pro Previewmedium79.0%50.52025-11-27
13Qwen3.5 397B A17Bthinking78.8%55.12026-02-16
14Gemini 3 Flashthinking78.0%59.12025-12-17
15Qwen3.7 Plus78.0%58.52026-06-01
16GPT-5.2-Codexxhigh77.6%59.42025-12-11
17Gemini 3.1 Flash Lite77.2%53.72026-03-03
18Gemini 3.1 Pro77.1%60.52026-02-19
19Qwen3.6 Maxmax76.6%62.32026-04-20
20DeepSeek V4 Pro 0424max76.5%59.62026-04-24
21Gemini 3.5 Flashhigh76.3%63.12026-05-19
22GLM 5.1thinking76.3%56.62026-04-07
23Kimi K2.6thinking76.0%59.42026-04-20
24Muse Spark75.9%64.32026-04-08
25GPT-5.4 Nanoxhigh75.9%54.22026-03-17
26GPT-5.5xhigh75.8%66.12026-04-23
27MiniMax M2.775.7%55.62026-03-18
28Qwen3.5-122B-A10Bthinking75.7%52.62026-02-24
29Qwen3.5-27Bthinking75.6%54.72026-02-24
30Gemma 4 31Bthinking75.6%52.02026-04-02
31GPT-5.2xhigh75.4%60.72025-12-11
32GPT-5 minihigh75.4%54.12025-08-07
33GPT-5.3-Codexxhigh75.4%63.82026-02-05
34Qwen3.6 Plus75.2%58.82026-04-02
35Gemini 3.5 Flashmedium74.6%62.82026-05-19
36GPT-5-Codexhigh74.2%57.52025-09-23
37GPT-5.4xhigh74.0%64.72026-03-05
38Command A+74.0%50.52026-05-20
39Gemma 4 12Bthinking73.5%50.52026-06-03
40DeepSeek V4 Flash 0420high73.5%55.42026-04-24
41GLM 5.2max73.3%62.32026-06-16
42GPT-5.4 Minixhigh73.3%56.32026-03-17
43GLM 5 Turbo73.2%55.52026-03-15
44GPT-5high73.1%57.92025-08-07
45GPT-5.1high72.9%57.72025-11-13
46GPT-5.6 Solmax72.7%68.02026-07-09
47Qwen3.5-35B-A3Bthinking72.5%52.82026-02-24
48Gemma 4 26B A4Bthinking72.5%50.02026-04-02
49GLM-5thinking72.3%58.92026-02-11
50MiniMax-M272.3%50.82025-10-26
51MiMo-V2-Flash71.8%51.62025-12-16
52GPT-5.5high71.6%65.32026-04-23
53MiniMax M2.571.6%53.62026-02-12
54GPT-5.5 Instant71.5%58.02026-05-05
55Nemotron 3 Super 120B A12Bthinking71.5%50.62026-03-11
56o371.4%59.52025-04-16
57DeepSeek V4 Pro 0424high71.3%57.82026-04-24
58GPT-5.6 Terramax71.2%64.42026-07-09
59Solar Pro 371.2%43.22026-04-06
60GPT-5 minimedium71.2%53.82025-08-07

More instruction following benchmarks · All LLM benchmarks

Cite as: BenchLeader, “IFBench leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/aa_ifbench.

IFBench: questions

What does IFBench measure?
IFBench tests precise instruction following with constraints the model has not seen before, such as unusual formatting or counting rules, applied to ordinary writing tasks. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads IFBench?
Grok 4.3 leads IFBench with 83.3% as of 25 Oct 2026, ahead of MiniMax M3 at 82.9%.
How many models have IFBench results?
423 model configurations have a IFBench result on BenchLeader, all taken from Artificial Analysis.
Who runs IFBench and how often is it updated?
IFBench is published by Artificial Analysis. BenchLeader re-reads the published results once a day and records the date each result was published.
Does IFBench count toward the BenchLeader Index?
Yes. IFBench contributes to the instruction following category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.