OpenAI: GPT-3.5 Turbo Instruct

Write a Python function that takes a list of dictionaries and returns only those where the 'status' key equals 'active'. Include type hints.

Here's a function that filters your list: ```python from typing import List, Dict, Any def filter_active(items: List[Dict[str, Any]]) -> List[Dict[str, Any]]: return [item for item in items if item.get('status') == 'active'] ``` This uses a list comprehension to iterate through each dictionary and checks if the 'status' key equals 'active'. The `.get()` method safely handles cases where 'status' might be missing.

The completion-based architecture excels at straightforward code generation with clear instructions. The 4095-token context window means this works well for isolated functions but struggles with large codebases requiring multi-file context. At $1.50/$2.00 per million tokens, it's cost-effective for high-volume simple tasks where newer chat models would be overkill.

Summarize this customer feedback in three bullet points, highlighting the main complaint and any positive mentions: [feedback about slow checkout, helpful staff, confusing return policy]

• **Main issue**: Checkout process is too slow, causing customer frustration during peak hours • **Positive note**: Staff members were helpful and patient in resolving questions • **Secondary concern**: Return policy documentation is unclear, leading to confusion at point of purchase The feedback suggests operational improvements in checkout flow and clearer policy communication would address most concerns.

The instruction-following format handles structured extraction tasks reliably. This model predates the conversational refinements in GPT-3.5 Turbo (chat), so it responds directly to imperative prompts without the chattiness of newer models. The limited context window means you'll need to batch long feedback threads rather than processing them in one pass.

Generate five creative tagline options for a sustainable coffee brand targeting urban millennials. Each should be under eight words.

1. Brew better. Planet first. 2. Morning ritual, zero guilt. 3. Conscious cups for city dwellers. 4. Sustainably sourced, seriously good coffee. 5. Wake up to a greener tomorrow. Each tagline balances environmental messaging with the premium quality expectation of the target demographic, using concise language that works across digital and print channels.

Creative generation remains a strength — the instruct model produces focused options without the over-explanation common in chat-tuned variants. However, it lacks the nuanced brand-voice calibration of GPT-4 class models. The output is competent but won't capture subtle tonal requirements without multiple refinement passes, which the shallow context window makes cumbersome.

When you need completion-style responses for older integrations

A 4-person dev shop maintaining a SaaS product built in 2020 has 18 customer-facing features that call OpenAI's legacy completion endpoint. GPT-3.5 Turbo Instruct is the only current model that preserves that completion format instead of chat messages. Migrating to chat-based models means rewriting prompt logic across the codebase and re-testing every integration. At $1.50/Mtok input, this model costs 40% less than GPT-4 variants while keeping your existing code intact. The 4095-token window handles most single-turn tasks like classification, extraction, and short-form generation. If your prompts regularly exceed 3000 tokens or you need multi-turn conversations, you're better off biting the migration cost now. Otherwise, this buys you 12-18 months before the completion API sunsets.

Cost-effective labeling when you process 500k+ items daily

A logistics company tags 600,000 customer support emails per day into 12 routing categories. Each email averages 180 tokens. GPT-3.5 Turbo Instruct runs this workload at roughly $270/day for input tokens versus $450+ on newer chat models with similar accuracy for simple classification. The completion format means less prompt overhead—no system message wrapping, just the text and the label list. Accuracy sits around 91% on their internal test set, which matches their threshold before human review. The 4095-token limit never binds because emails stay under 400 tokens. If classification quality drops below 88% or you add nuanced multi-label logic, upgrade to GPT-4o-mini. Until then, this model delivers the lowest per-unit cost in OpenAI's lineup for single-turn categorization at scale.

Cheapest OpenAI option for non-interactive bulk rewrites

A 9-person content agency rewrites 2,000 product descriptions monthly for an e-commerce client, turning technical specs into consumer-friendly copy. Each input is 300-600 tokens; output is similar length. They queue the job overnight in a Python script that hits the completion endpoint. At $2.00/Mtok output, the monthly bill is under $80 versus $120+ on chat-based models. The completion interface simplifies the script—no message array formatting, just prompt and max_tokens. Quality is good enough that 80% of outputs need only light edits. The 4095-token window accommodates their longest specs with room for examples. If the client demands brand-voice consistency across outputs or multi-step reasoning, switch to GPT-4o-mini and accept the 50% cost increase. For straightforward rewrite tasks where speed and interaction don't matter, this remains the most economical OpenAI model.