DeepSeek separates cache-hit, cache-miss and output tokens—and applies peak and off-peak rates. A time-of-day forecast that ignores the official UTC schedule can misprice a batch.
Pricing snapshot: Vendor first-party documentation checked on 25 October 2026. Original price effective date unknown unless separately declared. Amounts are published list figures, subject to tier, region, currency and eligibility.
The exact model behind the alias
DeepSeek’s official API pricing document identifies the public alias deepseek-flash as serving DeepSeek-V4.1-Flash-1017. It also says older flash-related aliases are accepted but now routed to the replacement. Preserve the requested alias and, when available, the resolved model version; a retired alias does not mean the original weights or original tariff are still running.
The same page lists a separate deepseek-v4-pro tier. Those price rows must not be merged simply because both APIs speak compatible chat formats. The 1017 suffix is a version identifier, not evidence of a billing effective date.
The flash rate matrix
For deepseek-flash, the checked source quotes off-peak prices per million tokens of $0.003 for cache-hit input, $0.15 for cache-miss input and $0.60 for output. Peak prices are double: $0.006, $0.30 and $1.20 respectively. Model cost therefore depends on output, how much input actually hit the cache and the provider-defined time window. A generic “input tokens × one price” equation cannot represent the product accurately.
These are USD list prices. They do not imply identical charges through every cloud marketplace or reselling platform.
UTC schedule, not local evening time
The vendor defines peak as 01:00–04:00 and 06:00–10:00 UTC on Monday through Friday, excluding Chinese public holidays. Other times, weekends and those holidays are off-peak under the published rule. Calculators should convert the execution timestamp into UTC and evaluate each usage record separately. A workload described simply as “run it after work in Istanbul” is too ambiguous to choose an off-peak rate reliably.
Use the actual execution or metering timestamp recognised by the provider; submitted-at and completed-at times may differ for queued work. Check the official exclusion calendar when a public holiday boundary matters.
Illustrative bill for two time windows
Imagine an off-peak job with 5 million cache-hit input tokens, 1 million uncached input and 1 million output tokens. Its model subtotal is 5 × $0.003 + 1 × $0.15 + 1 × $0.60 = $0.765. If the identical metered work ran wholly at the peak tariff, the subtotal would be $1.53. These numbers are a hypothetical load, not a promised saving or an assumption that every job is cacheable.
A scheduler may split a long job across peak boundaries. Segment actual usage into time-stamped billable records rather than pricing the whole monthly run at its start time.
Tools and hidden operational expenses
The same official document publishes separate Search and Fetch rates at $0.75 per thousand successful requests and announces a future change for server-side tool usage beginning 1 November 2026 in Beijing time. That future charge must not be described as already applicable on the 25 October cutoff. Keep a scheduled commercial change in a separate future-effective record, rather than backdating it.
Retries, context growth, database operations and verification work can exceed a cheap text model’s charges. The source also lists concurrency limits but does not guarantee immediate capacity for any particular organisation.
Evidence and practical implementation
Store rate IDs for cache hit/miss, output, peak/off-peak and separate tool calls; add a function to choose a rate only after the UTC schedule is known. Effective_at remains null for the token rate rows because the dynamic page does not identify the original introduction date. Availability is “listed,” not a guarantee of account entitlement. This pricing-specific article makes no independent claim about model quality or security of downstream deployments.
Pricing evidence boundary
This cost discussion concerns DeepSeek (deepseek-flash (DeepSeek-V4.1-Flash-1017)). Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- api-docs.deepseek.com — (official documentation; accessed 2026-10-25)