Skip to main content
Radar

Five or fewer confirmed AI changes, with original sources, on mornings when something changed.A free source-linked morning brief.

See the free Radar Brief
Instruction Following benchmark report

Best LLMs for Instruction FollowingSeptember 2026 Leaderboard

Data refreshed:

Ability to follow precise instructions and constraints

As of September 2026, the top instruction following model on the BenchLM leaderboard is MAI-Thinking-1 with a weighted instruction following score of 94.7.

Decision lens: use provisional-ranked mode for broader public evidence and verified-ranked mode for source-only comparisons. A model can move between views as evidence coverage changes.

Data refreshed
September 4, 2026
Provisional-ranked
120 of 417 models
Verified-ranked
124 of 417 models
Weighted evidence
1 of 3 benchmarks
3 tracked benchmarks

IFEval, IFBench, SOB Value Acc

Best Instruction Following picks

BenchLM summaries for instruction following plus the practical tradeoffs users check next: open weights, price, speed, latency, and context.

How BenchLM scores these

IFEval Leaderboard

Primary score: weighted instruction following score. Higher values rank first. Use the Show metric control to change the value shown in each row.

Updated Embed leaderboard

Switch between provisional-ranked and verified-ranked modes to compare the broader public dataset with sourced-only rankings.

Filters
Provisional-ranked mode includes source-unverified non-generated benchmark evidence.P = provisional benchmark row
Rank / modelWeighted Instruction Following
1
MAI-Thinking-1Microsoft · Closed
94.7%
2
Grok 4.3xAI · Closed
94.4%
3
GPT-5.2-CodexOpenAI · Closed
93.5%
4
MiniMax M3MiniMax · Open weight
93.5%
5
GLM-5.1Z.AI · Open weight
93.5%
6
MiMo-V2.5-ProXiaomi · Closed
93.5%
7
GPT-5.5OpenAI · Closed
92.9%
8
Muse SparkMeta · Closed
92.9%
9
GPT-5.4 nanoOpenAI · Closed
92.9%
10
MiniMax M2.7MiniMax · Open weight
92.7%
11
Qwen3.5-122B-A10BAlibaba · Open weight
92.7%
12
Gemma 4 31BGoogle · Open weight
92.6%
13
Qwen3.5-27BAlibaba · Open weight
92.6%
14
GPT-5.2OpenAI · Closed
92.3%
15
GPT-5.3 CodexOpenAI · Closed
92.3%
16
Qwen3.7 MaxAlibaba · Closed
91.1%
17
Qwen3.7 PlusAlibaba · Closed
91.1%
18
Qwen3.8 MaxAlibaba · Open weight
90.7%
19
GPT-5.4OpenAI · Closed
90.4%
20
Command A+Cohere · Open weight
90.4%
21
Gemma 4 12BGoogle · Open weight
89.8%
22
GLM-5.2Z.AI · Open weight
89.6%
23
GPT-5.4 miniOpenAI · Closed
89.6%
24
Inkling-SmallThinking Machines Lab · Open weight
89.6%
25
GLM-5-TurboZ.AI · Closed
89.4%

Top AI Models for Instruction FollowingSeptember 2026

As of September 2026, MAI-Thinking-1 leads the provisional instruction following leaderboard with a score of 94.7%, followed by Grok 4.3 (94.4%) and GPT-5.2-Codex (93.5%). BenchLM is currently showing 120 provisional-ranked models and 124 verified-ranked models in this category.

What changed

MAI-Thinking-1 leads the current instruction-following table on its sourced weighted row.

GPT-5.4 close second on IFEval and IFBench.

Claude Opus 4.6 holds #3, with strong IFBench scores.

Top models by benchmark

Instruction-following benchmark used in first-party comparison charts for agent-oriented reasoning models.(70% of category score)

RankModelReported score

Score in Context

What these scores mean

Instruction following carries a 5% weight in overall scoring — small, but it directly measures reliability. A model that ignores formatting rules, word count limits, or inclusion constraints is unusable in automated pipelines. The weighted score blends IFEval (verifiable constraints) and IFBench.

Known limitations

IFEval tests a specific set of verifiable constraints (word count, formatting, inclusion/exclusion) — it doesn't capture the full range of instruction-following quality. A model can score well on IFEval but still misinterpret nuanced or ambiguous instructions. IFBench is newer and coverage is still building.

How we weight

Instruction following carries a 5% weight in BenchLM.ai's overall scoring. While a smaller category, it directly measures reliability — a model that ignores constraints is unusable in production pipelines. See the instruction following leaderboard.

Leaderboards exclude benchmark rows that BenchLM generated from other scores or cloned from reference models. When a weighted benchmark is missing after that filter, the category falls back to the remaining trustworthy public rows instead of filling the gap with synthetic values.

The full scoring rules, freshness handling, and runtime/pricing caveats live on the BenchLM methodology page.

Scroll horizontally to read the full evidence ledger.

Instruction Following benchmark weights, ranking status, and descriptions
BenchmarkWeightStatusDescription
IFEvalDisplay onlyTests ability to follow verifiable instructions like format constraints and content requirements
IFBench70%WeightedInstruction-following benchmark used in first-party comparison charts for agent-oriented reasoning models.
SOB Value AccDisplay onlyA structured-output benchmark from Interfaze measuring whether extracted JSON leaf values exactly match verified ground truth.

About Instruction Following Benchmarks

Tests ability to follow verifiable instructions like format constraints and content requirements

Common questions

What is the best LLM for instruction following?

The best instruction-following LLMs are ranked by benchmarks like IFEval, which measure how accurately models adhere to specific formatting, length, and content constraints.

What is IFEval and how does it work?

IFEval (Instruction Following Evaluation) tests whether LLMs can precisely follow verifiable instructions such as word count limits, formatting rules, and inclusion/exclusion constraints.

Why is instruction following important for LLMs?

Instruction following is critical because real-world applications require models to reliably adhere to user specifications, output formats, and constraints rather than just generating plausible text.

Instruction following updates

Instruction following scores just got interesting. Get the weekly update.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.

Related