Checked August 29, 2026: DeepSeek V4 has two first-party API routes: Flash for inexpensive volume work and Pro for harder coding, reasoning, and analysis. Both now use peak and off-peak billing, so request time affects the bill as well as model and token volume.
The live table, chart, and calculator above read from AI Pricing Guru’s maintained dataset, refreshed August 29. They use the off-peak rate card as the baseline. Weekday peak calls cost twice the displayed input, cached-input, and output amount.
For the active provider table, open DeepSeek pricing. Use the token calculator for your traffic mix and compare DeepSeek vs OpenAI pricing before migrating production traffic.
DeepSeek V4 Flash vs V4 Pro
Start with V4 Flash when the task is high-volume, easy to verify, or safe to escalate. Strong candidates include classification, metadata extraction, query rewriting, short summaries, support drafts, RAG preprocessing, routing, and first-pass coding checks.
Choose V4 Pro when Flash fails too often or the task needs more judgment. It is the better DeepSeek candidate for repository changes, nuanced support, document analysis, multi-step reasoning, and second-pass review.
The useful metric is cost per accepted result. A cheaper route loses its advantage if it creates retries, escalations, slow reviews, or customer-facing errors. Test both models on the same representative cases and record first-pass success, output length, latency, and retry rate.
How current peak and off-peak billing works
Peak windows are 01:00–04:00 UTC and 06:00–10:00 UTC, Monday through Friday. All other times—including weekends—are off-peak.
DeepSeek describes off-peak as 50% below peak. In budget terms, peak input, cache-hit input, and output rates are each twice the off-peak baseline shown above. The calculator therefore gives the off-peak result; double its DeepSeek total when every token lands inside a peak window, or split mixed traffic by UTC period.
Batch workloads can respond by shifting queues. Interactive products should forecast against the peak tier because user requests cannot always wait. For the complete announcement analysis, read our DeepSeek pricing update.
Cached input and agent workloads
DeepSeek applies lower cache-hit rates to repeated prefixes. That can materially reduce the cost of systems that reuse stable system prompts, policy text, tool schemas, repository maps, retrieval templates, or long conversation scaffolds.
Do not assume the headline cache rate applies to most input. Track actual cached and uncached token counts in production. Small changes to a prefix can reduce hit rate, while dynamic context placed too early in the prompt can prevent reuse.
Cache-hit pricing follows the same time-of-day multiplier. Replay realistic long-prefix requests at both tiers before approving a budget; a short chat benchmark will not reveal the impact on a cache-heavy coding agent.
Best DeepSeek route by workload
| Workload | Start with | Escalate when |
|---|---|---|
| Classification and routing | V4 Flash | Labels fail validation |
| Structured extraction | V4 Flash | Retries erase the savings |
| Support drafts | V4 Flash | Tone or policy judgment is weak |
| RAG answers | V4 Flash | Evidence needs deeper synthesis |
| Mechanical code tasks | V4 Pro | Repository-wide judgment is required |
| Coding review | V4 Pro | Risk or ambiguity calls for a premium route |
| Regulated workflows | Approved vendor route | Governance requirements allow broader testing |
A layered architecture usually beats choosing one model for every step: route easy work to Flash, escalate uncertain cases to Pro, and reserve a premium provider for the small set of requests where vendor controls or task quality justify it.
Hidden costs to measure
Token price is only one part of the bill. Track retries, tool calls, output length, cache-hit rate, human review, latency, and provider fallback traffic. For agent systems, cap loop depth and tool use before comparing models; otherwise a low-cost call can trigger an expensive sequence.
Operational requirements matter too. Procurement, data handling, service commitments, regional availability, and incident support can outweigh a lower token rate. Compare current OpenAI pricing and Anthropic pricing when those controls are part of the decision.
For managed access to multiple open-model routes, check Novita’s current catalog and test the same workload against the direct DeepSeek API.
Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored link above at no extra cost to you. It does not affect our model recommendations.
Practical migration checklist
- Export token usage by UTC hour and weekday, then map traffic against both peak windows.
- Replay a representative evaluation set on Flash and Pro.
- Record cached input, uncached input, output, retries, and accepted results separately.
- Move delay-tolerant jobs to off-peak queues, including weekends when practical.
- Forecast live traffic at peak rates and treat off-peak execution as savings.
- Keep a tested fallback for high-value or time-sensitive requests.
DeepSeek remains a strong cost candidate, but time-of-day billing makes workload timing part of model economics. Use Flash as the volume layer, Pro as the harder-task tier, and measure the whole path to an accepted result.
Sources: DeepSeek’s official pricing documentation and the live AI Pricing Guru API. Pricing and model status verified August 29, 2026.