GPT-5.2 vs GPT-5 Mini vs GPT-5 Nano: 6 Constraint Tests
GPT-5.2 vs GPT-5 Mini vs GPT-5 Nano comes down to one question: which model follows constraints with the least babysitting. These six tests focus on machine-readable output, short summaries, and developer-style prompts.
Models tested
Test setup
- reasoning: low
- webSearch: false
- verbosity: low
Each test uses the same user prompt across models. System instructions only enforce output format (JSON only, SQL only, no extra text).
Test 1: Strict JSON (recipe schema)
Goal: return valid JSON with the requested schema and no extra keys.
Return ONLY valid JSON. Schema: {"recipe":{"name":string,"ingredients":[string],"steps":[string]},"notes":[string]}. Create a simple cold coffee recipe that uses oat milk and cinnamon.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
{"recipe":{"name":"Simple Oat Milk Cinnamon Cold Coffee","ingredients":["1 cup cold brewed coffee or strongly brewed coffee, chilled","3/4 cup oat milk","1-2 tsp maple syrup or sugar (optional)","1/4 tsp ground cinnamon, plus more for topping","1/2 cup ice","Pinch of salt (optional)"],"steps":["Add the coffee, oat milk, cinnamon, sweetener (if using), and salt (if using) to a glass or shaker.","Stir or shake until well combined.","Fill a serving glass with ice.","Pour the coffee mixture over the ice.","Sprinkle a little extra cinnamon on top and serve."]},"notes":["For stronger flavor, use cold brew concentrate and dilute with oat milk to taste.","If using hot coffee, chill it fully before mixing to avoid melting the ice too fast.","Cinnamon can clump; shaking in a sealed jar helps mix it evenly."]}
|
{
"recipe": {
"name": "Iced Oat Milk Coffee with Cinnamon",
"ingredients": [
"1 cup cold brewed coffee or strongly chilled brewed coffee",
"1/2 cup oat milk",
"1/2 teaspoon ground cinnamon",
"1-2 teaspoons maple syrup or sweetener of choice (optional)",
"Ice cubes",
"Cinnamon stick or pinch of cinnamon for garnish (optional)"
],
"steps": [
"Fill a glass with ice cubes.",
"Pour cold brewed coffee into the glass.",
"Add oat milk and stir gently to combine.",
"Stir in ground cinnamon and maple syrup until dissolved.",
"Taste and adjust sweetness or cinnamon as desired.",
"Garnish with a cinnamon stick or a light sprinkle of cinnamon and serve."
]
},
"notes": [
"Use cold brew for a smoother, less acidic flavor; chilled brewed coffee works fine.",
"Shake coffee, oat milk, cinnamon, and sweetener in a jar with ice for a frothier drink.",
"Adjust oat milk ratio for creaminess; more oat milk yields a milder coffee flavor.",
"This recipe is vegan if using a plant-based sweetener."
]
}
|
{"recipe":{"name":"Iced Cinnamon Oat Milk Coffee","ingredients":["1 cup cold brew or chilled coffee","1/2 cup oat milk","1-2 tsp sweetener (optional)","1/4 tsp ground cinnamon","ice cubes","1/4 tsp vanilla extract (optional)"],"steps":["Fill a glass with ice.","Pour chilled coffee over the ice.","Add oat milk and vanilla extract (if using).","Stir in ground cinnamon and sweetener until well mixed.","Serve immediately and enjoy."]},"notes":["Cinnamon can be adjusted to taste.","For a stronger cinnamon flavor, mix cinnamon with a little hot water to bloom before adding to the drink.","You can substitute cold brew for regular coffee if preferred."]}
|
All three stay inside the schema. GPT-5 Mini returns the most readable JSON, but the others stay valid.
Test 2: Two sentence summary with hard caps
Goal: exactly two sentences, no extra formatting.
Summarize the text below into exactly 2 sentences. Each sentence must be 18 words or fewer.
TEXT: Wiro is an API marketplace that lets developers try and ship AI models without vendor lock-in. It focuses on simple endpoints, predictable runtimes, and quick iteration across image, video, and audio models.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
Wiro is an API marketplace for developers to test and deploy AI models without vendor lock-in. It offers simple endpoints, predictable runtimes, and fast iteration across image, video, and audio models.
|
Wiro is an API marketplace letting developers try and ship AI models without vendor lock-in. It emphasizes simple endpoints, predictable runtimes, and rapid iteration for image, video, and audio models.
|
"Wiro is an API marketplace allowing developers to try and ship AI models without vendor lock-in." "It emphasizes simple endpoints, predictable runtimes, and rapid iteration across image, video, and audio models."
|
All models keep it to two sentences. GPT-5 Nano adds quotes around each sentence, which can break strict downstream parsing.
Test 3: SQL only output (top spenders)
Goal: return a working Postgres query with the requested output columns.
Write a Postgres SQL query. Return ONLY the SQL.
Schema:
- orders(id, user_id, total_cents, created_at)
- users(id, email)
Task: For the last 30 days, return the top 5 users by total spend. Output columns: email, total_spend_usd. Sort by total_spend_usd desc.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
SELECT u.email AS email, ROUND(SUM(o.total_cents) / 100.0, 2) AS total_spend_usd FROM orders o JOIN users u ON u.id = o.user_id WHERE o.created_at >= NOW() - INTERVAL '30 days' GROUP BY u.email ORDER BY total_spend_usd DESC LIMIT 5;
|
SELECT u.email, SUM(o.total_cents)::numeric/100 AS total_spend_usd FROM orders o JOIN users u ON u.id = o.user_id WHERE o.created_at >= now() - INTERVAL '30 days' GROUP BY u.email ORDER BY total_spend_usd DESC LIMIT 5;
|
SELECT u.email, SUM(o.total_cents) / 100.0 AS total_spend_usd FROM orders o JOIN users u ON o.user_id = u.id WHERE o.created_at >= NOW() - INTERVAL '30 days' GROUP BY u.email ORDER BY total_spend_usd DESC LIMIT 5;
|
All three produce usable SQL. GPT-5.2 adds a rounded USD value. The others return raw float/numeric values.
Test 4: JSON classification (modality grouping)
Goal: fill a fixed JSON schema and follow explicit rules.
Return ONLY valid JSON.
Task: Given this list of model names, group them by modality.
Models: ["openai/gpt-5.2", "openai/gpt-5-mini", "openai/gpt-5-nano", "klingai/kling-v3", "qwen/qwen3-asr-1-7b", "bytedance/seedream-v5-lite"]
Schema: {"text":[string],"image":[string],"video":[string],"audio":[string]}.
Rules: gpt models are text. klingai/kling-v3 is video. qwen/qwen3-asr-1-7b is audio. seedream is image.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
{"text":["openai/gpt-5.2","openai/gpt-5-mini","openai/gpt-5-nano"],"image":["bytedance/seedream-v5-lite"],"video":["klingai/kling-v3"],"audio":["qwen/qwen3-asr-1-7b"]}
|
{"text":["openai/gpt-5.2","openai/gpt-5-mini","openai/gpt-5-nano"],"image":["bytedance/seedream-v5-lite"],"video":["klingai/kling-v3"],"audio":["qwen/qwen3-asr-1-7b"]}
|
{"text":["openai/gpt-5.2","openai/gpt-5-mini","openai/gpt-5-nano"],"image":["bytedance/seedream-v5-lite"],"video":["klingai/kling-v3"],"audio":["qwen/qwen3-asr-1-7b"]}
|
This test comes out identical across the three models. The explicit rules help.
Test 5: Translation with name preservation
Goal: translate to Turkish without changing product names.
Translate the text below to Turkish. Keep product names unchanged. Return ONLY the translation.
Text: GPT-5.2 is tuned for agentic coding tasks. GPT-5 mini is faster for well-defined prompts. GPT-5 nano is best for lightweight classification.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
GPT-5.2 ajan tabanli kodlama gorevleri icin ayarlanmistir. GPT-5 mini iyi tanimlanmis istemler icin daha hizlidir. GPT-5 nano hafif siniflandirma icin en iyisidir.
|
GPT-5.2, otonom kodlama görevleri için ayarlanmıştır. GPT-5 mini, iyi tanımlanmış istemler için daha hızlıdır. GPT-5 nano, hafif sınıflandırma için en iyisidir.
|
GPT-5.2, ajan odaklı kodlama görevleri için ayarlanmıştır. GPT-5 mini, iyi tanımlanmış istemler için daha hızlıdır. GPT-5 nano, hafif sınıflandırma için en iyisidir.
|
All three keep product names unchanged. GPT-5.2 outputs a Turkish sentence without diacritics, while the others keep full Turkish characters.
Test 6: Tone control for support replies
Goal: rewrite an angry message into a calm reply under a word limit.
Rewrite the message below into a calm, professional support reply. Keep it under 70 words. Return ONLY the reply.
Message: Your upload failed again. This is the third time today. Fix your servers.
| GPT-5.2 |
GPT-5 Mini |
GPT-5 Nano |
Sorry for the trouble with your uploads. I understand this has failed three times today and that is frustrating. Please share the time of the attempts, any error message you saw, and the file type and size. In the meantime, try refreshing the page or signing out and back in. We are checking server performance and will follow up promptly.
|
Sorry you experienced repeated upload failures. We apologize for the inconvenience. Please provide the time(s) of failures, file type/size, and any error messages or screenshots. We will investigate our servers immediately and update you as soon as we have findings.
|
Thank you for reporting this. I apologize for the repeated upload failures today. Our team is investigating server issues and will update you shortly. Please try again with a stable connection and cleared cache. If the problem persists, reply with the file type, approximate time of attempt, and any error messages so we can assist quickly.
|
All three keep a calm tone and propose next steps. GPT-5 Mini is the most concise.
Runtime notes
These runs are short. The averages below come from task elapsedseconds across the six prompts.
| Model |
Avg seconds |
Min |
Max |
| GPT-5.2 |
6.0 |
4 |
8 |
| GPT-5 Mini |
7.8 |
7 |
11 |
| GPT-5 Nano |
7.5 |
4 |
10 |
Verdict
- Pick GPT-5.2 when outputs need strong structure and consistency across varied tasks.
- Pick GPT-5 Mini for format-heavy work where readable output matters (JSON, SQL, short replies).
- Pick GPT-5 Nano for simple summaries and classification, but watch for extra formatting like quotes.
Try it