OpenAI: GPT-4o (2024-05-13)

Review this Python function for potential bugs. It's supposed to merge two sorted lists but users report incorrect ordering in edge cases.

The function has a subtle off-by-one error in the merge logic. When both lists have remaining elements, it correctly compares values, but after exhausting one list, it uses `extend()` on the remaining list without checking if indices are valid. If `i` or `j` equals the list length, you'll append nothing or skip elements. Replace the final `extend()` calls with explicit `while` loops that append `list1[i]` and `list2[j]` individually while incrementing indices. Also consider the case where one input list is empty—your current code will raise an IndexError on the first comparison.

GPT-4o excels at systematic code review, tracing execution paths to find logic errors that pass surface-level inspection. The 128k context window means you can paste entire modules for holistic analysis. However, without benchmark data, we can't quantify its accuracy against specialized code models—treat suggestions as high-confidence hypotheses requiring validation.

I have a photo of a handwritten recipe card from my grandmother. Extract the full recipe with ingredients and steps, preserving her original measurements.

The model would process the uploaded image and return structured text: "**Grandma's Lemon Bars** — Crust: 1 c. butter (softened), 2 c. flour, ½ c. powdered sugar. Mix and press into 9x13 pan. Bake 350°F for 15 min. Filling: 4 eggs (beaten), 2 c. sugar, ⅓ c. lemon juice, ¼ c. flour, ½ tsp baking powder. Pour over hot crust, bake 25 min until set. Dust with powdered sugar when cool." It would capture the cursive handwriting, abbreviations like "c." and "tsp", and maintain the original informal tone rather than standardizing to modern recipe format.

The multimodal capability handles real-world image-to-text tasks that pure LLMs can't touch. Particularly strong with mixed content—handwriting plus printed labels, or diagrams with annotations. At $5 per million input tokens, processing high-resolution images is economical for document workflows. The model interprets context (recognizing "c." as cups, not degrees Celsius) but may struggle with severely degraded or stylized handwriting.

Write a 200-word product description for a minimalist desk lamp targeting remote workers. Emphasize adjustability and eye comfort without sounding clinical.

The model would generate: "Your desk doesn't need another gadget—it needs better light. This lamp bends where you need it: a three-axis arm that holds position without creep, and a head that rotates 340° so you can aim light at your keyboard, notebook, or the wall behind your monitor for indirect glow. The LED array mimics north-facing daylight—5000K color temperature that keeps whites neutral and reduces the yellow fatigue of evening work sessions. Touch controls on the base let you dim in 10% increments; hold to switch between focus mode (full brightness, tight beam) and ambient mode (60% output, diffused). The weighted steel base won't tip when you adjust mid-call. Available in matte black or warm gray. No app, no subscription, no firmware updates—just light where you need it, exactly when you need it."

GPT-4o balances technical precision with consumer-friendly tone, a combination that's difficult to prompt from earlier models. It avoids the clinical voice of spec sheets while keeping claims verifiable. The 128k context window is overkill for this task, but means you could provide brand guidelines, competitor descriptions, and product specs in one prompt. The $15/Mtok output cost adds up quickly for high-volume content generation—consider caching reusable context.

When GPT-4o handles screenshot-heavy support tickets at scale

A 12-person SaaS support team receives 200+ tickets daily, half with screenshots of UI bugs or config errors. GPT-4o's native image understanding lets you route tickets without manual tagging: it reads the screenshot, extracts error codes, checks against your knowledge base (via file upload), and drafts a response or escalation note. At $5 input / $15 output per Mtok, a 500-token ticket with one image costs roughly $0.01 to process. The 128k context window means you can include your entire product docs in the system prompt, so routing accuracy stays high without fine-tuning. If your ticket volume exceeds 500/day and images are rare, consider a text-only model to cut costs by 60%. Below that threshold, GPT-4o's multi-modal flexibility beats the operational overhead of splitting workflows.

Why GPT-4o works for full-contract review without chunking

A 4-attorney firm reviews 30-page vendor contracts (roughly 25k tokens) and needs to flag non-standard clauses against a master template. GPT-4o's 128k window fits the entire contract plus your 8k-token playbook in one call, so you avoid the chunking errors that plague smaller models. Output cost is the trade-off: at $15/Mtok, a 2k-token redline summary costs $0.03 per contract. If you're reviewing fewer than 50 contracts monthly, that's negligible. Above 200/month, the output bill ($6+ per contract if summaries run long) starts to hurt, and you should test Gemini 1.5 Pro at $3.75 output or Claude 3.5 Sonnet at $15 but with tighter output. For low-to-mid volume contract work where accuracy trumps cost, GPT-4o's context and reliability justify the premium.

When GPT-4o's speed and language range beat specialized models

A 20-person community platform moderates live chat in English, Spanish, and Mandarin, flagging harassment and off-topic spam in under 2 seconds. GPT-4o handles all three languages in one model (no language detection layer needed) and returns structured JSON flags fast enough for real-time use. At 1,000 messages/hour during peak, input cost is roughly $0.50/hour assuming 200 tokens per message. The lack of public benchmarks on this release means you're flying blind on safety recall rates compared to Llama Guard or Perspective API, so plan a 2-week shadow-mode test before going live. If moderation errors are high-stakes (regulated community, legal exposure), wait for benchmark data or use a specialized safety model as a second-pass filter. For general community moderation where speed and multilingual coverage matter more than perfect recall, GPT-4o is the pragmatic default.