DeepSeek V4.1 Flash API costs: a reproducible evaluation budget
Estimate DeepSeek V4.1 Flash evaluation costs with explicit token assumptions, time tiers, cache scenarios, and repeatable test cases.
Summary · TL;DR
A reproducible cost estimate for a fixed DeepSeek V4.1 Flash evaluation workload, with pricing assumptions, cache arithmetic, execution limits, and test cases.
Key Takeaways
- At the listed OneMux peak rates, the illustrative 150,000-request workload costs $24.219 if the catalogue amounts are confirmed to be per million tokens and no input is cached.
- The separately listed Fireworks AI US router rates produce a $310.50 estimate for the same 30-day workload, excluding cache and other charges.
- A repeatable evaluation needs fixed cases, output limits, retry rules, token-usage records, and a stop condition—not just a price per token.
- Verify OneMux billing units, tier boundaries, cache eligibility, and current prices before using these estimates as a purchasing decision.
On this page (8)
The short answer
For an illustrative evaluation of 150,000 requests, each using 3,000 input tokens and producing 400 output tokens, the estimated token cost is $24.219 at OneMux’s listed peak rates or $12.1095 at its off-peak rates. Those figures assume the OneMux catalogue amounts are USD per million tokens, no cached input, and no retries. Confirm the billing units and current rates before committing a budget: the supplied catalogue snapshot does not explicitly identify its token-price denominator.
For the same workload, the separately listed Fireworks AI US router rates yield $310.50, using the explicitly stated per-million-token units. These are price estimates for differently described routes, not a measured quality or performance comparison.
OneMux is the publisher and product discussed in this cost guide. The workflow below is useful regardless of which route you evaluate.
Price snapshot and comparable units
The OneMux catalogue snapshot was captured on November 12, 2026. The table reproduces its published amounts and ratio without changing their precision. For the calculations below, the OneMux amounts are treated as dollars per million tokens; that denominator needs confirmation on the model page or with the provider.
| OneMux catalogue tier | Official input | Official output | OneMux input | OneMux output | Published price ratio |
|---|---|---|---|---|---|
| Peak | $0.3 | $1.2 | $0.0351 | $0.1404 | 0.11699999999999999 |
| Off-peak | $0.15 | $0.6 | $0.01755 | $0.0702 | 0.11699999999999999 |
The published ratio is the OneMux price divided by the corresponding official price, not the percentage discount. Peak is listed as Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC; other times are off-peak. Confirm how a request crossing a boundary is billed.
For context, Future AGI’s Fireworks AI US router listing states $0.450 per million input tokens and $1.80 per million output tokens, with cached input at $0.009 per million. Its stated verification date is October 22, 2026. The listing identifies a particular Fireworks router; do not assume it is operationally identical to the OneMux route. Prices and terms can change.
Calculate a fixed workload
Use this illustrative workload for every route being compared:
- 100 fixed test cases, each repeated 50 times per day: 5,000 requests per day.
- A 30-day evaluation: 150,000 requests.
- Average billed usage of 3,000 input tokens and 400 output tokens per request.
- No cached input, failed attempts, tool-call rounds, taxes, or separate service charges in the base estimate.
Multiply requests by tokens per request, divide each token total by one million, then multiply by the applicable input or output rate. Add the two results.
| Workload component | Calculation | Result |
|---|---|---|
| Input tokens | 150,000 × 3,000 | 450 million |
| Output tokens | 150,000 × 400 | 60 million |
| OneMux peak, assuming per-million units | 450 × $0.0351 + 60 × $0.1404 | $24.219 |
| OneMux off-peak, assuming per-million units | 450 × $0.01755 + 60 × $0.0702 | $12.1095 |
| Fireworks US router | 450 × $0.450 + 60 × $1.80 | $310.50 |
These are not quotes. In particular, a maximum output setting is a ceiling, not a prediction of average billed output. If actual output averages 800 rather than 400 tokens, the output component doubles. For the OneMux peak scenario, that changes the total from $24.219 to $32.643, assuming all other inputs remain fixed.
A mixed-time run needs separate token totals for peak and off-peak requests. Apply each tier to the usage assigned to it rather than applying the off-peak price to an entire run merely because it started off-peak.
Model cache savings separately
The OneMux snapshot lists cache-read prices of $0.000702 peak and $0.000351 off-peak; their billing denominator also needs confirmation. The Fireworks listing explicitly states its cache-read price per million tokens. Cache eligibility, minimums, writes, and expiration are not established by the supplied material, so an eligible hit rate is an assumption, not a guarantee.
For example, assume a peak OneMux request has 3,000 input tokens, of which 2,400 are billed as cache reads and 600 as ordinary input, plus 400 output tokens. Under the per-million-unit assumption:
| Component | Calculation | Cost per request |
|---|---|---|
| Ordinary input | 600 ÷ 1,000,000 × $0.0351 | $0.00002106 |
| Cache reads | 2,400 ÷ 1,000,000 × $0.000702 | $0.0000016848 |
| Output | 400 ÷ 1,000,000 × $0.1404 | $0.00005616 |
| Total | Sum of the three components | $0.0000789048 |
That is $11.83572 for 150,000 identical requests, before any cache-write or other charges. The no-cache baseline remains essential: it is the safer estimate until actual cache-read usage and billing are observed.
Design a repeatable evaluation
Before running paid calls, freeze the case set, prompt templates, route, model identifier, and scoring rubric. The following cases are illustrative; they are not measured results for this model.
| Test case | Fixed input and check | Limit or failure condition |
|---|---|---|
| Structured extraction | Extract named fields from a short document; compare with a checked answer key. | Mark invalid or missing fields as failures; cap output tokens. |
| Long-context retrieval | Place a known fact in a document and ask for it with a supporting excerpt. | Record actual input size; reject unsupported or missing excerpts. |
| Tool-call selection | Present a request with one appropriate function and known arguments. | Check function name and arguments; cap tool-call rounds. |
| Refusal or uncertainty | Ask an unanswerable question based on the supplied material. | Flag invented facts; use the same rubric for every repetition. |
Set a maximum request count and a dollar stop threshold before execution. Limit output tokens, concurrency, retries per case, and tool rounds. Treat retries as additional billable attempts unless your provider confirms otherwise; log them separately rather than silently replacing failed calls. If the budget or error threshold is reached, stop and report incomplete coverage.
For each attempt, record the case ID, time in UTC, route and model identifier, input and output usage, cache-read usage if reported, result, and any error. Preserve the prompt and scoring version. Compare cost alongside pass rate and failure types, not in isolation. No paid evaluation was run for this article.
Limits and decision checks
The Fireworks listing reports a 1,048,576-token context window, a maximum of 393,216 output tokens, function calling, vision, reasoning, and prompt caching for its US router. These are documented listing claims, not measured performance. Its benchmark section reports no logged public benchmark scores. The OneMux catalogue describes context as 1M; it does not establish that every router-specific capability or limit is identical on OneMux.
Before buying, confirm the exact endpoint’s supported parameters, combined input/output limits, token accounting, cache terms, and current rates. If a capability is necessary for a test, run a small capped probe rather than assuming compatibility. Keep headroom below context and output ceilings, and investigate truncated responses, missing usage records, or unexpected cache misses before scaling.
A useful decision is therefore conditional: use the estimate to set a ceiling, then replace assumed token counts and cache behavior with observed usage from a bounded pilot.
Sources
- Future AGI: DeepSeek V4.1 Flash, Fireworks AI router, US — router prices, limits, and listed capabilities; source states verification on October 22, 2026.
- OneMux: DeepSeek V4.1 Flash — model catalogue and tier information; snapshot used here captured November 12, 2026.
Published price snapshot
Captured 2026-11-12T21:00:05.902Z 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.
| Model | Official input | Official output | OneMux input | OneMux output | Published price ratio |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.3 | $1.2 | $0.0351 | $0.1404 | 0.11699999999999999 |
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 a DeepSeek V4.1 Flash evaluation cost?
For 150,000 requests averaging 3,000 input and 400 output tokens, the illustrative OneMux peak estimate is $24.219, assuming its listed amounts are per million tokens and there are no cache reads or extra charges. Verify the billing denominator and current rates first.
Does prompt caching always reduce the bill?
Not necessarily. Savings depend on eligible cache reads, actual hit rates, and any applicable write or other charges. Calculate a no-cache baseline and verify billed usage.
Are the Fireworks router and OneMux routes proven to perform the same?
No. The supplied sources provide pricing and some route-specific capability information, not a controlled performance comparison. Test the exact routes and parameters you intend to use.
Related articles
Guides
Hy3 API cost estimates for repeatable evaluations
Estimate Hy3 evaluation costs with explicit token limits, cache assumptions, test cases and reproducible calculations.
Guides
DeepSeek V4.1 Flash 1017: Estimate API Costs Before an Evaluation
A reproducible guide to estimating DeepSeek V4.1 Flash 1017 API costs, with peak and off-peak rates, cache assumptions, and an evaluation budget.
Guides
GPT 6.2 Sol API cost: a reproducible evaluation budget
Estimate GPT 6.2 Sol evaluation costs with explicit token assumptions, context-tier pricing, cache examples, and repeatable test limits.
Guides
GPT 6.2 Astra API cost estimate for support triage with human approval
Estimate GPT 6.2 Astra token costs for a human-approved customer support triage workflow, with tier and cache examples.