GPT-5 Mini: 6 Practical Text Generation Tests
GPT-5 Mini targets fast, low-friction text generation. This review runs six small tests that show what the model returns when prompts demand strict formats and tight constraints.
Model link
Test setup
- webSearch: false
- verbosity: low
- reasoning: minimal, low, and medium (varies per test)
Test 1: Strict JSON output (recipe schema)
This test checks whether the model can return valid JSON with nested objects and arrays.
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.{"recipe":{"name":"Iced Cinnamon Oat Milk Coffee","ingredients":["1 cup strong brewed coffee, cooled","1/2–3/4 cup oat milk (to taste)","1–2 teaspoons simple syrup or sweetener of choice","1/4 teaspoon ground cinnamon plus extra for garnish","ice cubes","optional: whipped oat creamer or splash of vanilla extract"] ,"steps":["Brew strong coffee and let it cool to room temperature or chill in fridge.","Fill a glass with ice cubes.","Pour cooled coffee over the ice.","Add oat milk and sweetener, then stir to combine.","Stir in 1/4 teaspoon ground cinnamon until evenly mixed.","Taste and adjust sweetness or oat milk, add a splash of vanilla if desired.","Sprinkle a pinch of cinnamon on top and serve immediately." ]},"notes":["Use cold-brewed coffee for smoother flavor and less acidity.","For a frothier drink, shake coffee, oat milk, cinnamon, and sweetener with ice in a sealed jar before pouring.","Adjust cinnamon to taste; whole cinnamon stick can be used as a stirrer for extra aroma."]}The output stays inside the schema and does not add extra keys. That makes it easy to parse.
Test 2: Two sentence summary with hard limits
This test checks constraint following with sentence count and length caps.
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.Wiro is an API marketplace enabling developers to try and ship AI models without vendor lock-in. It offers simple endpoints, predictable runtimes, and fast iteration for image, video, and audio models.The summary follows the two sentence rule and keeps key points. It reads clean.
Test 3: SQL only response (top spenders)
This test checks whether the model can output code without explanation and match column requirements.
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.SELECT u.email, ROUND(SUM(o.total_cents)::numeric / 100, 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.id, u.email ORDER BY SUM(o.total_cents) DESC LIMIT 5;The query matches the schema and returns the requested columns. Formatting can be applied later.
Test 4: JSON classification (modality grouping)
This test checks simple classification into a fixed JSON schema. It also catches small but common mistakes.
Return ONLY valid JSON.
Task: Given this list of model names, group them by modality.
Models: ["openai/gpt-5-mini", "bytedance/seedream-v5-lite", "klingai/kling-v3", "qwen/qwen3-asr-1-7b"]
Schema: {"text":[string],"image":[string],"video":[string],"audio":[string]}.{"text":["openai/gpt-5-mini","klingai/kling-v3"],"image":["bytedance/seedream-v5-lite"],"video":[],"audio":["qwen/qwen3-asr-1-7b"]}This run mislabels klingai/kling-v3 as text and leaves video empty. The schema holds, but the classification needs a guardrail.
Test 5: Translation with name preservation
This test checks whether product names stay unchanged during translation.
Translate the text below to Turkish. Keep product names unchanged. Return ONLY the translation.
Text: Seedance Pro v1.5 generates short videos from text prompts and can also animate a starting image.Seedance Pro v1.5, metin istemlerinden kısa videolar oluşturur ve ayrıca bir başlangıç görüntüsünü canlandırabilir.The product name stays intact. The sentence reads natural.
Test 6: Tone control for support replies
This test checks tone shifts under a word limit.
Rewrite the message below into a calm, professional support reply. Keep it under 80 words. Return ONLY the reply.
Message: Your upload failed again. This is the third time today. Fix your servers.Sorry you experienced repeated upload failures. I apologize for the inconvenience. Please tell me the file name, file size, browser and version, time(s) of the attempts, and any error messages or screenshots. We will investigate server logs and escalate to engineering if needed. Meanwhile try clearing your cache, switching browsers, or using a different network. Reply with the details and I will open a priority ticket.The reply stays calm and asks for actionable debug details. It also suggests basic client-side retries while logs are checked.
Where GPT-5 Mini fits
- Format-heavy generation where outputs must be machine-readable
- Short summaries and rewrites with tight constraints
- Quick code and query drafts that can be validated by tests
