uogent®Analysis / Enterprise AI & Economics
What does a useful AI answer actually cost?
Measure AI spending against accepted work. Explore model charges, deployment costs and human review with three interactive calculators.
An AI system can turn a PDF into a brief in seconds. The usefulness of the brief is determined only after someone has validated its figures, qualifications and/or sources. That effort must be captured in the cost of the answer, just as much as the model call that produced it.
We use an example familiar to most businesses: attach a PDF and ask for its key facts, table figures, page references and uncertainties. The document might be a customer report, an operating update or a technical document. We want a brief a colleague can use, with enough evidence to verify the important claims.
In the example below, the Astra model charge is US$0.35 for a 10-page PDF. Once we include the platform, human checking and work that does not meet the required standard, the cost becomes about US$1.95 per useful brief. Following that difference is the purpose of this article.
We use one recurring example: a 10-page PDF, a two-page output budget and 1,000 documents a month. The worked examples use Astra. Each calculator also offers Fable, Sol and Opus, using direct API prices checked on 29 October 2026. The costs and review outcomes are planning assumptions, not measured model performance. All amounts are US dollars.
Defining a useful result
If attempting to determine the cost of your AI workload for the first time, define what the output must do. For this article, a useful output would preserve the important numbers and their units, flag exceptions, point to the relevant pages and say when the source is unclear. A fluent paragraph that misses a qualification would fail this test. The deadline matters too: a correct answer that arrives too late may not be usable. [1]
Use one PDF as the unit of work throughout. Several model requests may be needed to complete it, and they still belong in that one document’s cost. Include the time spent checking failed attempts.
Start with the model bill
Uploading a PDF can send both extracted text and page images to the model, allowing it to read tables, diagrams and layout. The charge depends on the content processed and the response generated. Providers (OpenAI, Anthropic etc) measure these in tokens: small units of text or other content. You can use pages in the calculators; the token estimates are indicative and vary among models and processing methods used. [11] [12] [13]
The main presets are 5-, 10- and 20-page PDFs, with outputs of one, two or five pages. Our rounded allowance is 3,000 input tokens per PDF page and 500 billed tokens per output page. The default therefore uses about 30,000 input tokens and 1,000 output tokens. Actual usage varies by file and model.
Output pages are a budget estimate. Billed reasoning can use some of that allowance, leaving less visible text. Formatting and language also affect the conversion. These are business API costs; they are not a separate charge for each PDF attached to a consumer subscription. [9] [10]
At the standard API rates used here, Astra charges US$0.30 for the input and US$0.05 for the output. That gives a model charge of US$0.35 for one document, or US$350 for 1,000 documents. [2] [4] [5] [6]
One 10-page PDF, a two-page output budget and one request.
With Astra selected and the two-page output unchanged, five input pages cost US$0.20 and twenty cost US$0.65. Then try the model selector: the default 10-page PDF costs US$0.35 with Astra or Fable, US$0.07 with Sol and US$0.14 with Opus. These are the same token allowances at different model prices; actual quality and token use must be tested for the. [17] [18]
How the page estimates and model charge are calculated
These are rounded planning allowances, not measured token counts or provider upload limits. For a 10-page PDF, we allow 5,000 text tokens, 20,000 image tokens and 2,000 instruction tokens. We add 3,000 tokens of headroom to reach 30,000. Headroom is a reserve in the estimate, not an additional provider fee. Tables and images are already included in the page allowance.
One output page represents a budget of 500 billed tokens, roughly 375 words of ordinary English before allowing for reasoning. The default two-page budget is 1,000 tokens. It does not guarantee two visible pages. [11]
For the Astra worked example within the main range, the standard rates are US$10 per million input tokens and US$50 per million output tokens. Dividing by one million converts those published rates to a single request.
A dense scan may use more tokens than a clean digital report. The same PDF can also consume different amounts on different models. Replace these estimates with usage measured from representative uploads before using them to set a budget. [14] [15] [16]
Add the cost of making the answer useful
The model bill is the first layer. Add the platform used to run the workflow and the time spent checking its answers. Then divide the full operating cost by the number of briefs that met the agreed standard. Costs from unsuccessful attempts belong in the total too. [3]
Continue with the same 1,000 PDFs. The model bill is US$350 and the platform allowance is US$300 a month. Each document takes three minutes to check at a labour cost of US$24 per hour, including employment overheads. At US$0.40 a minute, that is US$1.20 per document, or US$1,200 for the month.
Table 1 · One monthly workflow
Item | Monthly amount or result |
|---|---|
Model charges | US$350 |
Platform allowance | US$300 |
Human review | US$1,200 |
Total operating cost | US$1,850 |
Briefs meeting the standard | 950 of 1,000 |
Cost per useful brief | US$1.95 |
We assume 950 briefs meet the standard after any checking and correction included in that review time. Fifty remain unresolved. The US$1,850 total covers attempts and review across all 1,000 documents, so dividing by 950 captures the cost of producing work the business can use.
Start with review time. Increase it from three to six minutes, holding everything else steady: the cost per useful brief rises from US$1.95 to US$3.21. Next reduce acceptance from 95% to 80%: the result becomes US$3.81. More checking and fewer useful outputs both increase the cost. These changes illustrate the arithmetic; they are not predictions about any of the four models.
The 50 unresolved documents still matter. Their attempted processing and checking are already included, but later manual completion is not. If they must be finished, add that effort and update the completion count. When no briefs meet the standard, the calculator shows the unit cost as undefined instead of dividing by zero.
This is the practical lesson: a low model price helps, but the time needed to reach a dependable answer can matter more. Before optimising the token bill, identify which part of the workflow creates avoidable effort.
When does private AI become worth considering?
Hosted AI means calling a commercial model through its API. The provider operates the model; the business still runs the application around it and pays for usage. Private AI means operating a separate model on owned or rented infrastructure. In this example, private does not mean installing these commercial models on your own server.
For the comparison below, keep the same 10-page PDF and two-page output. Hosted uses Astra, with a US$300 monthly platform allowance plus the US$0.35 model charge per document. Private uses an illustrative US$3,000 monthly commitment plus US$0.02 per document. Start at 1,000 documents, then increase the volume to see how those different cost structures behave.
The private budget is an assumption that covers document preparation, visual analysis or OCR where required, model operation, storage and support. It must also have enough capacity for the workload. The calculator does not automatically increase that budget when files become longer or requests multiply. Break-even points and analyses are directional and indicative.
Table 2 · Monthly technology costs for the same PDF workflow
Documents / month | Hosted | Private |
|---|---|---|
1,000 | US$650 | US$3,020 |
5,000 | US$2,050 | US$3,100 |
10,000 | US$3,800 | US$3,200 |
These figures exclude human review and setup, and assume sufficient private capacity. Orange shows hosted costs; the blue dotted line shows private. The chart starts at 0–10,000 documents so the crossover remains easy to see.
At 1,000 documents, hosted has the lower commitment. At 5,000, it still costs less: US$2,050 against US$3,100. By 10,000 documents, private is lower at US$3,200 against US$3,800. Under these assumptions, the crossover is around 8,200 documents a month. That demonstrates why private AI can become commercially meaningful as a recurring workload grows.
The cost difference is useful only if the private workflow can deliver the required quality and speed. Check review effort, peak demand and response times before treating it as a saving. Changing the model selector changes the hosted price; it does not establish equivalent private-model performance. Lower hosted rates can move the crossover beyond the initial 10,000-document view. Larger volumes remain available under Advanced assumptions.
How the crossover is calculated
Private commits an extra US$2,700 each month: US$3,000 minus US$300. Its assumed variable cost saves US$0.33 per document: US$0.35 minus US$0.02. Divide the extra commitment by that saving.
The first whole document at which private is strictly cheaper is 8,182. A larger private fixed budget moves the crossover later. A higher private cost per document reduces the saving available to recover it. If private has no variable-cost advantage, increasing volume alone does not create a crossover. Use Advanced assumptions to change these budgets.
Improve the workflow, then test the saving
Start with the largest avoidable cost. Clear page references may shorten checking. Better source files may reduce correction. Sending the relevant pages or requesting a shorter answer may lower model usage, provided important context is preserved. Recheck acceptance after each change: a saving that creates extra correction work may disappear elsewhere.
Optional: larger workloads, repeated requests and setup
Advanced assumptions keeps the larger scenarios available without making them the starting point. Enable larger workloads to use up to 100 input PDF pages, 20 output pages and, where relevant, 100,000 documents a month. The normal view allows up to 20 input pages, five output pages and 10,000 documents a month. Each calculator is independent; its profile line shows the current workload. Reset restores the 10-page, two-page-output example and, where applicable, 1,000 documents.
A larger upload can cross a pricing boundary. At our allowance of 3,000 input tokens per page, Astra and Sol cross their 272,000-input-token threshold at 91 PDF pages. Above it, the whole request uses twice the input and cache rates and 1.5 times the output rate. Fable and Opus retain their standard rates across the supported range. The calculators apply these differences automatically, including Batch and cached input. Exact rates appear below each result. [2] [5] [17] [18]
Repeated requests incur more processing. The 10-page reuse example makes ten requests with the same PDF content and a two-page output budget on each. It reuses 25,000 tokens of PDF text and images; instructions and headroom remain uncached. One write and nine successful reads within five minutes cost about US$1.54 for Astra or US$1.37 for Fable, compared with US$3.50 uncached. [7] [8]
Cache savings require matching, unchanged content at the start of the input and compatible rendering. The first cache write replaces the ordinary input charge for that portion; it is not added on top. Reused PDF pages cannot exceed input pages. Inspect usage records to confirm that later requests actually benefit.
Batch can reduce eligible model rates by 50% when the work can wait. In the default Astra monthly workflow, that saves US$175 on a US$350 model bill, about 9% of the full US$1,850 operating cost. Check that the turnaround meets the deadline. [2] [4]
Include setup allocates the illustrative US$12,000 upfront cost over 12 months, adding US$1,000 per month and taking the default cost to US$3.00 per useful brief. This allocation supports comparison; the cash payment remains upfront. Model, platform, review and capacity assumptions should be checked before using the results for an investment decision.
Compare hosted and private using the same acceptance rules, even if their underlying models or document-handling steps differ. A lower technology bill is valuable when the workflow can deliver the required work reliably. Include the people needed to operate it, the time to resolve failures and the capacity needed during busy periods.
Time released creates capacity before it creates cash savings. Record whether that capacity clears a backlog, improves service or avoids spending that would otherwise occur. At sufficient volume, private AI can be commercially attractive. The deciding factor is the cost of dependable, accepted work, supported by a workflow the business can actually operate.
Sources and calculation notes
Prices checked 29 October 2026. Quogent scenarios are illustrative. Direct API prices exclude taxes, regional premiums, tools and negotiated terms. Displayed prices are rounded to cents. See the expandable notes and calculator workings for detailed assumptions.
[1] Anthropic · Evaluation and success criteria
[3] FinOps Foundation · Unit economics
[6] Anthropic · Claude Fable 5.1
[8] Anthropic · Prompt caching
[10] Anthropic · Billed thinking
[14] OpenAI · Image token accounting
[15] Anthropic · Vision token accounting