benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $2.50 |
| Standard output | $15.00 |
| Cached input | $0.25 |
| Long-context (input >272K) | surcharge applies |
| Context window | up to 1,000,000 tokens |
| Max output | 128,000 tokens |
The model id is gpt-5.4. The family launched March 5, 2026 as GPT-5.4 Thinking and GPT-5.4 Pro, with mini and nano following on March 17. Mini is the free-tier ChatGPT model; nano is API-only. GPT-5.5 took the flagship crown seven weeks later, which is exactly why this page exists: superseded models get price-stable, and price-stable mid-tiers are where the value hides.
Where it sits: the OpenAI ladder
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| GPT-5 Mini | $0.25 | $2.00 | 400K |
| GPT-5 | $1.25 | $10.00 | 400K |
| GPT-5.4 | $2.50 | $15.00 | up to 1M |
| GPT-5.5 | $5.00 | $30.00 | 1.05M |
Read the ladder by what each doubling buys. GPT-5 to GPT-5.4: the context ceiling jumps from 400K to 1M, computer use arrives, and the finance work gets a tuned model. That's a real capability step. GPT-5.4 to GPT-5.5: the benchmark ceiling rises (84.0% SWE-bench Verified, officially published) and you pay double again for it. Teams that need the context but not the absolute ceiling are the ones the middle tier is for. The OpenAI pricing guide walks the whole lineup.
The 272K long-context surcharge
Like GPT-5.5, GPT-5.4's advertised million-token context comes with a pricing cliff: inputs beyond the 272K-token standard window bill at higher rates per OpenAI's pricing page. The failure mode is a long-running session that quietly crosses the line and reprices the whole input. The defenses are the same ones the 272K-cliff explainer covers for 5.5: track session token counts, summarize-and-restart before the line, or chunk via retrieval. If your sessions never approach 272K, the cliff is irrelevant and the headline rate is what you pay.
Cached input at $0.25
GPT-5.4's cached input costs $0.25/1M: less than GPT-5's standard input and one fifth of Sonnet 4.6's cached rate. An 80K-token stable prefix called 5,000 times a day bills roughly $100/day at the cached rate, compared with $1,000 uncached.
That example covers the repeated input only. Run your full input, output, and cache mix in the cost calculator.
The Excel angle
GPT-5.4 launched alongside ChatGPT for Excel, tuned on real finance workflows — modeling, scenario analysis, long-form research. OpenAI reported its internal investment-banking benchmark jumping from 43.7% with GPT-5 to 87.3% with GPT-5.4 Thinking, and OSWorld-Verified computer use at 75% against a 72.4% human baseline. Those numbers are OpenAI's own, but they explain the model's positioning: it was built to be the professional-work tier, not the everything tier. Spreadsheet-heavy teams should read the spreadsheets roundup before picking.
Use-case fit
Consider for: Financial modeling and document analysis; computer-use pipelines that do not need GPT-5.5's ceiling; context-heavy work in the 400K–1M range; and prompt-heavy agents that can use the $0.25 cached rate.
Skip if: You want OpenAI's strongest — that's GPT-5.5 at double the price; your tasks fit in 400K context and short prompts — GPT-5 at half price covers it; you need cheap volume — GPT-5 Mini costs a tenth.