Best LLMs for Instruction Following — OpenAI
The best OpenAI model for instruction following is GPT-5.2-Codex, satisfying 77.6% of IFBench's verifiable prompt constraints. GPT-5.4 Nano (75.9%) and GPT-5.5 (75.9%) round out the top three.
- 1
gpt-5.2-codexReasoningToolsJSON+140.1 intel · $1.75/M · 400K ctx77.6%IFBench - 2gpt-5.4-nanoReasoningToolsJSON+138.2 intel · $0.200/M · 400K ctx75.9%IFBench
- 3gpt-5.5ReasoningToolsJSON+154.8 intel · $5.00/M · 1.1M ctx75.9%IFBench
- 4gpt-5.2ReasoningToolsJSON+142.2 intel · $1.75/M · 74 t/s75.4%IFBench
- 5gpt-5-miniReasoningToolsJSON+125.3 intel · $0.250/M · 84 t/s75.4%IFBench
- 6gpt-5.3-codexReasoningToolsJSON+144.3 intel · $1.75/M · 122 t/s75.4%IFBench
- 7gpt-5.4ReasoningToolsJSON+151.4 intel · $2.50/M · 1.1M ctx73.9%IFBench
- 8gpt-5.4-miniReasoningToolsJSON+140.0 intel · $0.750/M · 400K ctx73.3%IFBench
- 9gpt-5ReasoningToolsJSON+134.7 intel · $1.25/M · 65 t/s73.1%IFBench
- 10gpt-5.1ReasoningToolsJSON+136.9 intel · $1.25/M · 51 t/s72.9%IFBench
- 11gpt-5.6-solReasoningToolsJSON+158.9 intel · $5.00/M · 50 t/s72.7%IFBench
- 12o3ReasoningToolsJSON+130.4 intel · $2.00/M · 44 t/s71.4%IFBench
- 13gpt-5.6-terraReasoningToolsJSON+155.0 intel · $1.00/M · 60 t/s71.2%IFBench
- 14o1ReasoningToolsJSON+123.4 intel · $15.00/M · 200K ctx70.3%IFBench
- 15gpt-5.1-codexReasoningToolsJSON+134.7 intel · $1.25/M · 66 t/s70.0%IFBench
- 16gpt-oss-120bReasoningToolsJSON23.8 intel · $0.037/M · 490 t/s69.0%IFBench
- 17o4-mini-highReasoningToolsJSON+125.6 intel · $1.10/M · 95 t/s68.7%IFBench
- 18o4-miniReasoningToolsJSON+125.6 intel · $1.10/M · 81 t/s68.7%IFBench
- 19gpt-5.1-codex-miniReasoningToolsJSON+130.6 intel · $0.250/M · 51 t/s67.9%IFBench
- 20gpt-5-nanoReasoningToolsJSON+119.9 intel · $0.050/M · 81 t/s67.6%IFBench
- 21o3-mini-highReasoningToolsJSON15.6 intel · $1.10/M · 200K ctx67.1%IFBench
- 22gpt-oss-20bReasoningToolsJSON14.9 intel · $0.030/M · 274 t/s65.1%IFBench
- 23gpt-oss-20b:freeReasoningToolsJSON14.9 intel · Free/M · 17 t/s65.1%IFBench
- 24gpt-4.1ToolsJSONVision19.4 intel · $2.00/M · 1.0M ctx43.0%IFBench
- 25gpt-4.1-miniToolsJSONVision14.8 intel · $0.400/M · 1.0M ctx38.3%IFBench
How this is ranked
AI models ranked by IFBench — how reliably a model obeys explicit, verifiable constraints in the prompt (format, length, inclusion and exclusion rules). High scores mean fewer retries and less prompt-wrangling in production, which often matters more than raw intelligence for pipelines.
Frequently asked
Which OpenAI model follows instructions most reliably?
The best OpenAI model for instruction following is GPT-5.2-Codex, satisfying 77.6% of IFBench's verifiable prompt constraints. GPT-5.4 Nano (75.9%) and GPT-5.5 (75.9%) round out the top three.
What's a good alternative to GPT-5.2-Codex?
GPT-5.4 Nano (75.9%) is the closest alternative on this metric, followed by GPT-5.5 (75.9%). See the full ranking above for the tradeoffs.
How many OpenAI models are there?
modelgrep tracks 60 OpenAI models with live benchmarks, speed, latency and per-provider pricing, led on intelligence by GPT-5.6 Sol. 25 of them qualify for this ranking.