GuidesOneMux

DeepSeek V4.1 Flash 1017 API Costs for Document Question Answering

Estimate document question-answering costs using explicit token assumptions, published time-based prices, and a reproducible caching example.

AI

Summary · TL;DR

An illustrative document QA workload of 150,000 requests per 30-day month costs $207 at the cited $0.30 input and $1.20 output rates per million tokens. The corresponding OneMux catalogue estimate is $34.98 at peak rates or $17.49 off-peak, before caching and other workload costs.

Key Takeaways

  • At 3,000 input and 400 output tokens per request, the cited reference rate yields $0.00138 per request before caching.
  • The OneMux catalogue lists different peak and off-peak rates; a real monthly estimate must reflect when requests are billed.
  • Source labels, retrieved passages, and citation text consume tokens and should be included in workload measurements.
  • Caching can reduce input charges when eligible tokens are billed at the cache-read rate, but eligibility and actual hit volume must be verified.
  • Model token charges exclude retrieval infrastructure, document processing, retries, and other possible production costs.
On this page (9)

The estimate in brief

For an illustrative document question-answering workload of 5,000 requests a day, each with 3,000 input tokens and 400 output tokens, DeepSeek V4.1 Flash 1017 costs $207 over a 30-day month at the cited reference rates. The equivalent estimate using OneMux’s published peak rates is $34.98; using its off-peak rates for every request gives $17.49. These are model-token estimates, not complete application bills.

OneMux publishes this blog and is one of the products discussed below. Its catalogue lists time-dependent prices; the reference prices come from Future AGI’s Fireworks AI model page, which identifies LiteLLM as its pricing source. Check current provider terms and billing units before committing a budget.

Prices and units

The table uses USD per 1 million tokens, matching the external source’s stated unit. The OneMux catalogue snapshot was captured on November 12, 2026; confirm the current billing unit and rates on the model page before using the figures for purchasing. The external page says its figures were verified October 22, 2026.

Price basisOfficial inputOfficial outputOneMux inputOneMux outputPublished OneMux-to-official ratio
Peak$0.30$1.20$0.0507$0.20280.169
Off-peak$0.15$0.60$0.02535$0.10140.169

The ratio is the OneMux price divided by the corresponding official price, not the percentage discounted. A ratio of 0.169 corresponds to an 83.1% difference for these listed token rates. It does not establish the total savings for an application.

In the OneMux snapshot, peak is Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC; other times are off-peak. The external Fireworks AI page lists $0.30 input and $1.20 output without establishing these OneMux time conditions. Treat its rate as a separate reference, not proof of an identical provider-side schedule.

Build a document QA request budget

Assume an application retrieves passages, sends them with a question, and asks the model for an answer with source attribution. For this example, each request contains:

  • 3,000 input tokens: instructions, the question, retrieved passages, and passage labels.
  • 400 output tokens: the answer and its source references.
  • One model call per question: no cache reads, retries, or additional tool-call rounds.
  • 5,000 questions per day for 30 days: 150,000 requests per month.

These are planning assumptions, not measured performance. The monthly volume is 450 million input tokens and 60 million output tokens. Multiplying each by its rate per million produces the following:

Pricing scenarioInput costOutput costTotal per requestTotal per 30-day month
External reference rate$135.00$72.00$0.001380$207.00
OneMux peak for every request$22.815$12.168$0.00023322$34.983
OneMux off-peak for every request$11.4075$6.084$0.00011661$17.4915

For example, the reference-rate calculation is 450 × $0.30 plus 60 × $1.20, or $207. The monthly figures use exactly 30 days; longer months or different daily traffic change the result.

An actual OneMux bill may include requests in both time tiers. Split measured token volumes by their applicable tier, calculate input and output charges separately, and add the results. The two all-peak and all-off-peak rows are scenarios, not predictions of a real traffic mix. Confirm how the provider assigns a tier to requests near a boundary.

Account for source attribution

Attribution has a direct token cost: document titles, identifiers, page or section labels, and quoted evidence occupy input space; citations and quotations in the answer occupy output space. Include them in token measurements rather than budgeting only for the user’s question and answer prose.

A practical request can label retrieved passages with stable IDs such as policy-v3, section 4 and ask the model to cite those IDs beside supported claims. Keep the mapping from each ID to the underlying document in the application so a reader can inspect the source. If no supplied passage supports an answer, instruct the model to say so rather than inventing a citation.

Check citation quality separately from price: sample answers, verify that cited passages actually support their claims, and test questions whose answers are absent or spread across documents. A plausible-looking source label is not evidence that the claim is correct. The supplied source does not provide a document QA accuracy benchmark for this model.

When cached input changes the estimate

The external page lists cached input at $0.006 per million tokens. The OneMux snapshot lists cache-read prices of $0.001014 per million at peak and $0.000507 off-peak. These are cache-read rates, not a promise that every repeated document passage will qualify.

Consider a separate illustrative scenario: of each 3,000-token input, 2,400 tokens are eligible cache reads and the remaining 600 are ordinary input. Output remains 400 tokens; the workload remains 150,000 requests. Assume no separate cache-write or storage charge for this calculation, but check actual provider terms for any such charges.

ScenarioOrdinary input per monthCached input per monthOutput per monthMonthly token cost
External reference rates$27.00$2.16$72.00$101.16
OneMux peak rates$4.563$0.36504$12.168$17.09604

For the reference row, 90 million ordinary input tokens cost $27, 360 million cached tokens cost $2.16, and 60 million output tokens cost $72. This example is useful only if usage records confirm that proportion of input is actually billed as cache reads. Cache eligibility, matching rules, expiration, and potential additional charges must be checked with the service used.

Limits and costs outside the model call

The Future AGI model page reports a 1,048,576-token context window, up to 393,216 output tokens, and support for prompt caching and function calling. Those are documented specifications, not evidence that a long-context document workflow will be accurate or economical. PDF input is not advertised on that page; plan for a document extraction step unless your chosen service documents another supported path.

The estimates exclude extraction and OCR, embeddings, search or vector storage, application hosting, and human review. They also exclude extra model calls and retry tokens. If retrieval returns too much irrelevant material, input costs rise without necessarily improving attribution; if it returns too little, the model may lack evidence. Measure those tradeoffs on representative documents before scaling.

Conclusion

Start with measured input, output, cache-read, and time-tier volumes rather than a single blended token price. For the stated 30-day example, the cited reference estimate is $207, while the OneMux catalogue scenarios range from $17.49 to $34.98 before caching. Validate both citation quality and the full application bill before choosing a route.

Sources

Published price snapshot

Captured 2026-11-12T14:30:05.729Z from the OneMux published catalogue. Prices retain website billing units; this is not a live benchmark. Tiered, cached or time-based billing may differ. Unknown values are not zero.

ModelOfficial inputOfficial outputOneMux inputOneMux outputPublished price ratio
DeepSeek V4.1 Flash 1017$0.3$1.2$0.0507$0.20280.169

Current OneMux prices. Compare workload costs only when input/output and billing units match; see each model page for tier conditions.

FAQ

How much does DeepSeek V4.1 Flash 1017 cost for document question answering?

At the cited reference rates of $0.30 per million input tokens and $1.20 per million output tokens, a request with 3,000 input and 400 output tokens costs $0.00138 before caching or other charges. At 150,000 such requests, the estimate is $207.

Do citations increase document QA API costs?

They can. Passage labels and source metadata add input tokens, while citations and quotations in the answer add output tokens. Measure both in representative requests.

Does the cache-read example guarantee lower costs?

No. It assumes 2,400 of 3,000 input tokens per request qualify for cache-read pricing. Actual eligibility, cache usage, and any additional charges depend on the service and its terms.

Are the OneMux estimates the full cost of a document QA application?

No. They cover the stated model-token workload only. Document extraction, retrieval infrastructure, additional calls, retries, hosting, and review may add costs.

Related articles