OpenAI: GPT-3.5 Turbo (older v0613)

Summarise this 800-word customer support transcript into three bullet points highlighting the customer's main issue, our resolution, and any follow-up needed.

This example would produce a concise three-bullet summary extracting the core complaint (e.g. billing discrepancy), the support agent's resolution steps (refund issued, account credited), and next actions (customer to verify receipt within 3-5 days). The model structures information clearly and maintains factual accuracy when condensing straightforward dialogues, though it may occasionally miss nuanced subtext or emotional undertones present in the original exchange.

GPT-3.5 Turbo v0613 excels at structured summarisation tasks where the input follows predictable patterns. The 4K token window handles most support transcripts comfortably. However, this older checkpoint predates instruction-tuning improvements in later releases, so complex multi-step reasoning or ambiguous requests may produce less reliable outputs than current models.

Write a professional email declining a meeting request for next Tuesday. Keep it polite, suggest alternative times later in the month, and keep it under 100 words.

This example would generate a courteous 80-word email opening with appreciation for the invitation, clearly stating unavailability on Tuesday, and proposing two specific alternative slots (e.g. 'Would the 24th at 2pm or the 27th at 10am work for you?'). The tone remains professional without being stiff, and the structure follows standard business email conventions—greeting, decline, alternatives, closing.

Straightforward templated writing is a strength of this checkpoint. The model reliably follows length constraints and tone guidance for routine correspondence. The v0613 release handles basic instruction-following well, though it lacks the nuanced style adaptation and context retention of GPT-4 or later 3.5 versions when requests involve multiple conflicting constraints.

Extract all product names, prices, and SKU codes from this e-commerce page HTML snippet and return them as a JSON array.

This example would parse the HTML, identify product data within common e-commerce markup patterns (divs with class names like 'product-card', 'price', 'sku'), and return a clean JSON array: [{'name': 'Wireless Mouse', 'price': '$24.99', 'sku': 'WM-2301'}, ...]. The extraction handles standard formatting reliably, though irregular HTML structures or obfuscated class names may cause the model to miss entries or hallucinate fields not present in the source.

This checkpoint performs adequately on structured data extraction from semi-predictable formats like e-commerce HTML. The 4K context window limits how much page content you can process in one request—roughly 2-3 product listings with full markup. For large-scale scraping or complex nested structures, newer models with larger windows and better parsing offer fewer errors.

When GPT-3.5 Turbo v0613 makes sense for early-stage chat experiments

A 3-person startup building their first customer-facing chatbot should start here. At $1.50 blended per million tokens, this model costs roughly 20× less than GPT-4 variants while handling straightforward Q&A, appointment booking, and FAQ routing without issue. The 4095-token context is tight—you'll hit limits on conversations longer than 8-10 exchanges—but for prototyping flows and validating user intent patterns, that constraint forces good design. If your bot sees under 10,000 messages/month and doesn't need nuanced reasoning or multi-turn memory, this older checkpoint delivers functional responses at a price that won't drain seed funding. Once you prove the concept and traffic crosses 50,000 messages/month, migrate to a newer 16k-context model.

Why this model still works for simple support-ticket tagging at scale

A 12-person e-commerce support team routing 2,000 inbound emails daily can use GPT-3.5 Turbo v0613 to tag tickets into 6-8 categories (refund, shipping, product question) before human review. Each classification call uses roughly 150 tokens input and 10 tokens output—at $1.50 blended that's $0.00024 per email, or about $15/month for 60,000 tickets. The model handles single-label classification reliably when categories are distinct and the prompt is well-structured. You'll see accuracy drop on edge cases (complaints spanning multiple categories, sarcasm, non-English), but for 80% of volume this checkpoint is fast and cheap enough to justify the occasional misroute. If accuracy falls below 85% after spot-checking 200 tickets, upgrade to GPT-4o-mini for $0.15 blended and tighter reasoning.

When to use this model for internal stand-up notes under 1,000 words

A 6-person product team recording daily 15-minute stand-ups (roughly 800-word transcripts) can feed those transcripts to GPT-3.5 Turbo v0613 for bullet-point summaries without hitting context limits. At 1,200 tokens input and 200 tokens output per summary, you're spending $0.0016 per meeting—negligible even at 250 meetings/year. The model extracts action items and blockers competently when the transcript is clean and the team uses consistent terminology. It struggles with crosstalk, multiple speakers on the same topic, or meetings that reference prior context outside the current transcript. If your stand-ups run longer than 20 minutes or reference previous decisions frequently, the 4k window becomes a bottleneck and you'll need GPT-4o-mini's 128k context to avoid truncation.