OpenAI: GPT-4

Refactor this Python function to use list comprehensions and remove the nested loops. Keep the same logic but make it more Pythonic.

The model would produce a clean refactor that replaces nested for-loops with a single list comprehension or generator expression, preserving edge cases like empty inputs. It typically adds inline comments explaining non-obvious transformations and suggests type hints if the original code lacked them. The refactored version maintains readability while reducing line count by 30-40%.

GPT-4 excels at idiomatic rewrites that balance brevity with clarity. Its 8K context window handles medium-sized functions comfortably, though it may truncate explanations when refactoring larger modules. The model occasionally over-optimizes for conciseness at the expense of beginner readability.

Explain the CAP theorem to a product manager who needs to choose between PostgreSQL and Cassandra for a new feature. Focus on practical trade-offs, not theory.

The model would frame CAP theorem as a decision tree: PostgreSQL guarantees consistency and tolerates partition failures by sacrificing some availability (transactions may block). Cassandra prioritizes availability and partition tolerance, meaning reads might return stale data briefly. It would then map these to product scenarios—'use Postgres if you need strict inventory counts; use Cassandra if you need a global activity feed that can't go down.' The explanation avoids academic jargon.

GPT-4 translates technical concepts into business context effectively, a strength for cross-functional documentation. The 8K window allows it to include 2-3 concrete examples without truncation. However, it sometimes hedges excessively ('it depends') rather than making a clear recommendation when one exists.

Draft a three-paragraph email declining a vendor proposal. Tone: polite but firm. Reason: their pricing model doesn't align with our usage patterns, but we want to stay in touch.

The model would produce a structured email opening with appreciation for the proposal, a second paragraph diplomatically explaining the pricing mismatch (e.g., 'our intermittent usage makes flat-rate licensing cost-prohibitive'), and a closing paragraph expressing interest in future conversations if their model evolves. The tone balances professionalism with warmth—no corporate clichés, no over-apologizing.

GPT-4 handles nuanced communication tasks well, capturing tone constraints without sounding robotic. At $60/Mtok output, this use case is expensive relative to simpler models that could draft emails adequately. The model shines when the task requires reading subtext or navigating political sensitivity.

When GPT-4 justifies the premium on high-stakes writing

A 4-person consulting shop sends 8-12 proposals a month, each going through 3-4 revision cycles with client feedback. GPT-4 at $30/$60 per Mtok costs roughly $0.18 per 3,000-token proposal draft—negligible against a $40k contract. The 8k context window holds the full RFP, previous proposal sections, and client comments in one prompt, so the model rewrites with all context in view. Cheaper models at this context length either hallucinate requirements or lose thread between sections. The cost threshold: if you're drafting under 20 documents a month where each mistake costs you a deal, GPT-4's reliability pays for itself. Beyond that volume, test whether GPT-4o or Claude 3.5 Sonnet hit your quality bar at half the output cost.

Where GPT-4 loses to newer models on repetitive extraction

A 10-person finance team processes 200 invoices a week, pulling vendor names, line items, and totals into a database. GPT-4 can handle the task but at $30 input per Mtok, scanning 200 PDFs (average 4k tokens each) costs $24/week or $1,248/year. The 8k context cap means multi-page invoices need chunking, which introduces errors at page boundaries. GPT-4o runs the same workload at $2.50 input per Mtok—$2/week—with a 128k window that swallows entire documents. GPT-4's instruction-following was best-in-class in 2023, but for high-volume structured tasks where the schema is fixed, newer models with lower input pricing and larger windows are the correct call. Use GPT-4 here only if you're on a legacy integration that hasn't migrated to the newer endpoint.

Why GPT-4's latency and cost don't fit live chat workflows

A 15-person SaaS startup wants an AI to read incoming Slack support threads and tag them as billing, technical, or sales before routing. GPT-4 handles the classification but averages 4-6 seconds per response and costs $0.24 per 1,000 customer messages at typical token counts. Over 500 messages a day, that's $120/month and a noticeable lag that frustrates customers expecting instant acknowledgment. GPT-4o delivers sub-2-second responses at $15/month for the same volume, and the 128k context window lets the model see the full thread history without truncation. GPT-4 made sense when it was the only reliable classifier; now it's the wrong tool for any workflow where humans are waiting on the model to respond. Save it for the complex escalations that need the extra reasoning depth.