An enterprise OpenAI API budget should include the selected model’s billable input and output, applicable cache categories, processing mode, and any separately charged tools. Forecast those quantities for your workload before approving spend. If you buy through Token360, use your enterprise quote rather than treating OpenAI’s public rates as your payable prices.
This article answers a specific purchasing question: what does an OpenAI API rate card mean for a monthly budget? It provides a dated price snapshot and transparent arithmetic, not a benchmark or a quote. For the broader process of comparing suppliers, use our LLM API quote comparison guide.
What are OpenAI’s public API prices?
OpenAI publishes model-specific rates rather than one universal API price. The snapshot below covers three text models at Standard, short-context pricing, in US dollars per million tokens. It is a public-list-price reference checked on November 20, 2026; it does not describe a negotiated Token360 enterprise offer.
Selected OpenAI public rates · USD per 1 million tokens · November 20, 2026
| Model | Uncached input | Output | Cached input | 30-minute cache write | 12-hour cache write |
| gpt-6.2-sol | $2.00 | $10.00 | $0.10 | $2.50 | $5.00 |
| gpt-6.2-astra | $8.00 | $40.00 | $0.40 | $10.00 | $20.00 |
| gpt-6.1-sol | $2.00 | $10.00 | $0.10 | $2.50 | $5.00 |
Source: OpenAI API pricing, checked November 20, 2026. This is a selected snapshot, not the complete model catalog. Confirm the context range and service mode before using any row.
For example, the GPT-6.2 Sol model documentation says requests above 272,000 input tokens use higher rates for the full request: input and cache rates double, while output costs 1.5 times the short-context rate. It also distinguishes Standard from faster processing modes. A long-document workflow therefore needs its own forecast rather than borrowing a short-chat estimate. Source checked November 20, 2026.
For procurement, record the exact model ID and rate-card date alongside the approved budget. “OpenAI model” is not a sufficiently precise line item: the table shows that choosing a different model can change both sides of the calculation. Model quality and suitability still need evaluation; a price table alone cannot select the model for your business.
What would 100,000 requests cost in a month?
In a simplified GPT-6.2 Sol example, 100,000 requests with 2,000 uncached input tokens and 500 billable output tokens each cost $900 at the public Standard short-context rates above. That is a hypothetical token subtotal, not an observed invoice, an all-in project budget, or a Token360 quote.
The assumptions deliberately exclude cache reads and writes, hosted tools, retries, other models, taxes, and any additional contractual charges. The 500-token assumption means billable output, not a word count or an estimate of visible answer length. Actual usage must replace these assumptions before purchase approval.
Illustrative monthly arithmetic, not a customer result
| Component | Monthly quantity | Calculation | Subtotal |
| Uncached input | 200 million tokens | 200 × $2 | $400 |
| Billable output | 50 million tokens | 50 × $10 | $500 |
| Text-token subtotal | 100,000 requests | $400 + $500 | $900 |
The rate inputs are independently shown on the GPT-6.2 Sol model page. The workload quantities are editorial assumptions. Their separation matters: a reader can update the rates without mistaking the workload for a published customer measurement.
Two sensitivity checks show what the budget depends on. If request count doubles while usage per request stays constant, the same subtotal becomes $1,800. If request count stays at 100,000 but billable output doubles to 1,000 tokens per request, it becomes $1,400. These are arithmetic scenarios under unchanged rates, not predictions of how a deployed application will behave.
Choose the scenario that represents the business change. Adding customers changes volume; generating longer reports changes usage per request. Do not approve a “per request” price based on short answers and then apply it unchanged to a report-writing workflow.
Which charges can sit outside the text-token subtotal?
A text-only subtotal is incomplete when the application uses separately billed services. Identify the actual tools, storage, and modalities in the proposed workflow, then budget their published meters separately. Do not assume every API call has an extra tool fee, or that one input-token rate covers all of these services.
OpenAI’s pricing page lists web search at $10 per 1,000 tool calls, plus search-content tokens at the model’s rates. It lists file-search storage at $0.10 per GB per day, with the first GB free, and file-search tool calls at $2.50 per 1,000 calls for the Responses API. These are distinct meters. See OpenAI’s tool pricing, checked November 20, 2026.
For illustration, 10,000 web-search tool calls would add $100 in call charges. They would not make the complete monthly total exactly $1,000: the $900 example did not include the search-content tokens or any other added usage. Keep that missing quantity visible instead of silently treating it as zero.
For an enterprise assistant, ask engineering whether “one user question” produces one model response, several model calls, a search, or document retrieval. Finance needs the resulting billable quantities. It does not need to reverse-engineer the application from a monthly total. If usage includes reasoning, the separate reasoning-token billing guide explains the accounting questions to resolve.
How should caching and Batch affect the forecast?
Caching and Batch should appear as separate, justified budget scenarios. Caching depends on which tokens are actually read from or written to a cache; Batch depends on whether the work can be processed asynchronously. Neither should be entered as an assumed discount across every request before the workload has been checked.
What should the caching scenario include?
OpenAI’s prompt-caching documentation distinguishes uncached input, cached input, and cache writes. Cache-write pricing is a replacement rate for the relevant input tokens, not an additional input charge on those same tokens. Record the applicable write lifetime and avoid counting the same tokens twice. Source checked November 20, 2026.
For an early budget, show a conservative case without assumed cache savings and a separate case supported by measured usage categories. An instruction repeated in your application is not, by itself, evidence that the entire request qualifies for the cached-input rate. If the application has not been tested, label the caching benefit as unverified.
When is the Batch price relevant?
The OpenAI Batch API guide describes asynchronous processing with a 50% cost discount versus synchronous APIs and a completion window of up to 24 hours. Source checked November 20, 2026. That makes it a candidate for delay-tolerant work, not a substitute for the interactive service your users expect.
Separate nightly document processing from customer-facing answers in the budget. Before assigning Batch rates, verify that the selected model, endpoint, and request features are supported. A lower rate is useful only for the portion of work that can actually use that processing mode.
How does Token360’s enterprise tiered pricing change the budget?
Token360 uses tiered enterprise quotes: higher usage receives lower quoted rates. Token360 also does not charge a separate service fee. Enterprises should obtain the rates and eligible usage for their own workload; neither statement makes inference free or establishes a universal discount against OpenAI’s public price list.
The Token360 enterprise page describes custom pricing based on volume, model mix, and production requirements. The tiered-quote and no-separate-service-fee statements above were confirmed by Token360 for this article. Specific price bands, eligible models, and commercial conditions must be established in the enterprise offer.
For the workload in the example, ask for a Token360 estimate at the expected monthly usage and at a realistic growth level. Keep the model, input/output profile, and required service conditions identifiable in both. The useful result is a payable estimate at each volume, not an assumed percentage taken off the OpenAI table.
Two different pricing references
| Reference | What it establishes | What it does not establish |
| OpenAI public rate card | A dated public baseline for the named model and processing scope | The price of a Token360 route or a negotiated enterprise contract |
| Token360 enterprise quote | The offered rates for the specified volume, model mix, and requirements | Automatic discounts for all accounts or an unspecified universal savings percentage |
Clarify what happens when usage moves into another band: when the new rate applies, which usage counts toward the band, and whether the offer contains a minimum commitment. These are questions to settle in the quote, not claims that a particular billing mechanism is already in place. Our AI API volume-discount guide covers the different mechanisms in detail.
A lower unit rate can coexist with a higher total bill if usage grows enough. That is why the enterprise forecast should show both the applicable rate and the expected payable amount. The commercial benefit of a lower rate should not be confused with a promise that total spend falls as adoption increases.
What should finance approve before production?
Approve a dated monthly forecast with explicit assumptions, the selected supply route, and a process for replacing estimates with observed usage. For direct OpenAI use, that forecast starts with the applicable OpenAI rates. For Token360, it starts with the agreed enterprise rates and conditions for the same intended workload.
A practical approval note can be short:
- Scope: model ID, API operation, processing mode, and whether requests cross a long-context pricing threshold.
- Usage: expected monthly requests and billable input/output quantities, with cache and tool assumptions stated separately.
- Commercial basis: rate-card date or enterprise quote version, applicable volume band, and any confirmed minimum or exclusions.
- Review trigger: a material change in model, answer length, tool use, or monthly volume that requires a refreshed forecast.
After launch, compare measured usage and settled charges with the approved assumptions. Token360’s billing documentation explains that an upstream provider’s public price is not automatically the Token360 route price and that final charges depend on measured usage and the active pricing rule. Use the API billing reconciliation guide for the operational follow-through.
When requesting a Token360 enterprise quote, send the model IDs, expected monthly usage, input/output profile, and required service conditions. Ask for the applicable tiered rates and a worked monthly estimate. That gives procurement a concrete purchase to assess while keeping OpenAI’s public pricing in its proper role: a dated reference, not a substitute for the actual offer.
About this article: written by Yuki Yu for Token360. Public documentation was checked on November 20, 2026. Numerical workloads are hypothetical; no live spending test or negotiated savings measurement is claimed. Recheck rates and terms before committing spend.