Short answer: OpenRouter listed GPT-5.4 at $2.50 per million input and $15.00 per million output tokens on November 13, 2026, rising to $5.00/$22.50 once a prompt reaches 272,000 tokens. GPT-5.4 mini is $0.75/$4.50, nano is $0.20/$1.25, and pro is $30.00/$180.00 with its own override. Cache reads are $0.25 per million on the standard route and batch halves the base rates.
| Route | Prompt | Completion | Note and meter |
|---|---|---|---|
| openai/gpt-5.41,050,000 context | $2.50 | $15.00 | Override at 272,000 prompt tokens: $5.00/$22.50. |
| openai/gpt-5.4:batchbatch | $1.25 | $7.50 | Half the standard rate; the same override applies. |
| openai/gpt-5.4-mini400,000 context | $0.75 | $4.50 | A third of the standard input rate. |
| openai/gpt-5.4-nano400,000 context | $0.20 | $1.25 | Cheapest text route in this family. |
| openai/gpt-5.4-pro1,050,000 context | $30.00 | $180.00 | Override at 272,000 tokens: $60.00/$270.00. |
| openai/gpt-5.4-image-2272,000 context | $8.00 | $15.00 | Image route with a separate image-output meter. |
One threshold explains most GPT bill surprises
GPT-5.4 lists $2.50 input and $15.00 output below 272,000 prompt tokens and $5.00/$22.50 at or above it. The override doubles input and raises output by 50%. A retrieval pipeline that usually sends 150,000 tokens can still cross the line on long documents, and those requests are repriced rather than surcharged proportionally.
The fix is to measure the tail. Sort your real prompt sizes and compute what share exceeds the threshold. If 4% of requests cross it, price two buckets and add them: the base bucket at one rate, the crossing bucket at the other. If the crossing bucket contributes more than a few percent of spend, check whether retrieval depth can be bounded or a summarization step can shrink the prompt below the threshold more cheaply than the override costs.
Prompt composition also decides whether caching and the threshold behave as expected. A single long system prompt, a fixed tool catalog, and retrieved documents each contribute tokens, and only the stable portion is a realistic cache candidate. Log the components separately for a week: the system prompt is usually constant, retrieval varies per request, and conversation history grows until it is trimmed. That breakdown tells you which control actually reduces the crossing rate.
The family is a routing menu
Mini at $0.75/$4.50 and nano at $0.20/$1.25 are not just cheaper versions of the same product; they are different quality tiers. The useful pattern is a cascade: classify or extract with nano, escalate ambiguous cases to mini, and reserve the standard model for work that failed the cheaper tiers. The saving depends entirely on the escalation rate.
Compute the cascade before building it. If nano handles 60% of traffic at $0.20 input and the remaining 40% goes to GPT-5.4 at $2.50, the blended input rate is $1.12 per million — better than standard but worse than mini alone. If nano's quality gate fails and 90% escalates, the same design adds latency and complexity for almost no saving. A bounded evaluation with your own rubric is the only way to know.
Cache reads and batch on the standard route
The standard route lists a $0.25 cache-read rate against $2.50 input, a 90% reduction on cached tokens. That is meaningful for agent traffic with a stable tool-definition prefix, and it is why logging cache-read tokens matters: the discount only appears on the invoice if the provider recognised the prefix.
Batch halves the base rates to $1.25/$7.50 with the same override structure. As with every batch tier, it only applies to asynchronous work. Evaluation runs, bulk classification, and overnight document processing fit; a user waiting for a response does not. Combine batch with the threshold check before quoting a monthly figure, because a batch job with oversized prompts still pays the override.
- Cache reads at $0.25 against $2.50 input reward a stable prefix such as tool definitions.
- A cache write you cannot reuse is a premium, not a saving; measure the hit rate first.
- Batch and cache stack arithmetically in this calculator, but only if both conditions are true.
- Web search is a separate $0.01-per-call meter on this route and is not part of the token estimate.
Worked estimate: 50,000 long-document requests
Take 50,000 requests per month with 40,000 input tokens and 800 output tokens, 20% cache hits, all below the override. Uncached input is 1.6B tokens at $2.50 ($4,000.00), cached input is 400M at $0.25 ($100.00), and output is 40M at $15.00 ($600.00). Total $4,700.00 per month, and input is 87% of the bill.
Now move 10% of the traffic (5,000 requests) to 300,000 prompt tokens. That slice alone carries 1.5B input tokens — 75% of the original 2B — and every one of them bills at $5.00 input and $0.50 cached input instead of $2.50 and $0.25. Modelled with the same 20% cache hit rate and 800 output tokens, the slice costs $6,210 and the unchanged 45,000 requests cost $4,230, for $10,440 a month against the $4,700 baseline: a 122% increase from a tenth of the requests. Context control is the single highest-leverage lever in this workload, and it costs nothing to measure.
Reconciling against APIDir's accepted rows
APIDir's accepted observation for the GPT API on OpenAI is $0.20 input and $1.20 output with a 'standard under 270k context' tier, checked August 22, 2026. The route retrieved in November for GPT-5.4 is $2.50/$15.00. Those rows describe different offers, possibly different models within the family, and the older row predates the current catalog structure.
Do not merge them into one average. Use the row that matches the exact model identifier your client sends, keep the checked date visible, and recheck the provider page before a budget commitment. APIDir's publication rule is the same one this guide follows: a price without a date and a source is not evidence.
Budget a GPT-5.4 workload
Enter the base rates and your token mix. If a meaningful share of requests crosses 272,000 prompt tokens, price those separately at the override rates.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
How much does GPT-5.4 cost per million tokens?
OpenRouter listed $2.50 input and $15.00 output per million tokens on November 13, 2026, doubling to $5.00/$22.50 at or above 272,000 prompt tokens.
What triggers the GPT-5.4 price override?
The prompt size, not the output size. Once a request reaches 272,000 prompt tokens, the route's override rates apply to that request.
Is GPT-5.4 mini enough?
That depends on your task. Mini is $0.75 input and $4.50 output, a third of the standard input rate. Test it on your own evaluation set and measure the escalation rate before routing production traffic to it.
Does batch pricing halve GPT-5.4?
The batch route lists $1.25 input and $7.50 output, half the standard rates, and the same override structure applies above the threshold.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.