OpenAI GPT-4 Pricing

Current per-token rates

GPT-4o: $2.50/1M input, $10/1M output. GPT-4 Turbo: $10/$30. GPT-4o-mini: $0.15/$0.60. All rates billed per token, no minimums.

Cached input pricing

GPT-4o cached input is billed at 50% of standard ($1.25/1M). Cache hits trigger automatically on identical prefixes ≥1024 tokens.

Batch API discount

Async jobs submitted via Batch API run at 50% off both input and output. Completes within 24h. Ideal for evals, embeddings backfills, and bulk summarization.

Worked cost example

A support agent: 4K tokens in, 600 out per call, 100K calls/month. GPT-4o ≈ $1,600. GPT-4o-mini ≈ $96. Tier choice changes the bill by 16×.

When each tier fits

GPT-4o: reasoning, tool use, accuracy-critical paths. Turbo: legacy compatibility only. Mini: chat, classification, extraction, anything where mini benchmarks within 5% of flagship.

Mechanics: how the cache-hit threshold works

GPT-4o's automatic caching only triggers on prompt prefixes of at least 1,024 tokens that match a previous request exactly from the start. A system prompt shorter than that, or one with a dynamic value inserted at the beginning rather than the end, never qualifies — restructuring prompts so static content comes first is often the only change needed to unlock the 50% cached-input rate.

Decision checklist: GPT-4o vs GPT-4o mini

— Does the task require multi-step reasoning or tool orchestration? Lean flagship. — Is it classification, extraction, or short-form chat? Test mini first. — What's the cost delta at your actual volume — run both through the calculator. — Can you route by confidence score, falling back to flagship only on low-confidence mini output?

Trade-offs and failure modes in tier selection

Defaulting everything to GPT-4o is the most common overspend: at 16x the mini rate on this workload, a wrong default compounds fast at volume. The opposite failure — routing accuracy-critical calls to mini to save cost — risks silent quality degradation that's harder to detect than a billing spike. Neither failure shows up until you've modelled both tiers side by side.

How to model a tier-migration scenario

To test moving a workload from Turbo to GPT-4o, hold your input/output token volumes fixed and swap only the per-1M rates: Turbo at $10/$30 versus GPT-4o at $2.50/$10 on the same 100K-call, 4K/600-token workload swings the bill from roughly $4,720 to $1,600 — a 66% reduction from a rate change alone, with zero change in actual usage.

Frequently asked questions

How is GPT-4 billed?

Per token, with input and output priced separately and output priced higher. Cached input tokens are billed at a reduced rate.

Why is my GPT-4 bill higher than expected?

Usually output length or prompt growth. A system prompt that doubles or an unbounded response length multiplies cost across every request.

Should I use a smaller model instead?

Often, yes. Price the same workload on the mini tier — for classification, extraction and routing tasks the quality difference is frequently smaller than the cost difference.