The rate card
| Model | Input /1M | Output /1M | Cached read | Cache write | Batch input | Context |
|---|---|---|---|---|---|---|
| GPT-5.4 | $2.5 | $15 | $0.25 | — | $1.25 | 272K |
| GPT-5.4 mini | $0.75 | $4.5 | $0.075 | — | $0.375 | 272K |
What it costs on real workloads
Realistic monthly cost at 1,000 requests/day with the calculator’s defaults — 8% retries, 40% cache hit rate, 15% infra overhead, per-model tokenizer calibration applied. The same three workloads used across every comparison on this site, so the numbers line up.
| Workload | GPT-5.4 | GPT-5.4 mini |
|---|---|---|
| Simple chatbot | $258/mo | $77/mo |
| Chatbot with history | $490/mo | $147/mo |
| Multi-step agent | $3,231/mo | $969/mo |
Three things move the bill more than the rate card
1 · Cached input costs 10% of standard
Repeated prompt prefixes bill at $0.25/1M instead of $2.5/1M on the flagship — a 90% discount. Anything with a stable system prompt, a reused document, or a growing conversation transcript should be hitting this. On the GPT-5.4 line it applies automatically above a token threshold; you don’t mark cache breakpoints yourself.
2 · There is no cache-write premium
This is the quiet advantage of the OpenAI line and it doesn’t appear on any pricing page as a line item — because it’s the absence of one. Warming the cache costs standard input price. Some competitors bill first-time writes at a premium, which means at low hit rates you pay extra for a cache you never reuse. Here, a cache miss is just a normal call. Bursty, low-volume traffic is where that matters most.
3 · Batch is 50% off, if you can wait
Asynchronous batch runs at $1.25/1M input — 50% below real-time — at the cost of minutes-to-hours latency. Only useful for offline work: nightly summarisation, bulk classification, document pipelines. Never for anything a user is waiting on.
Flagship or mini?
On the history-carrying chatbot in the table above, GPT-5.4 runs $490/mo and GPT-5.4 mini $147/mo — mini is 70% cheaper for the identical workload. That is a far larger lever than any provider-switching decision, so it deserves to be the first question rather than an afterthought.
Mini is a genuine step down in reasoning, not a free lunch — it is well suited to classification, extraction, summarisation and routing, and poorly suited to multi-step reasoning or code generation where a wrong answer costs a retry (and the retry costs more than the tier saved). If ultra-low cost is the constraint, OpenAI also publishes a nano tier below mini that sits outside this catalog.
o200k_base, which is the baseline our calculator calibrates every other model against — so token counts you measure here transfer directly, with no ratio adjustment.Run your own numbers on the GPT-5.4 line — your tokens, traffic, cache hit rate, and retries.
Open the calculator →