Is a subscription always cheaper? No. Low or irregular usage, unused allowances and cached input can favor pay-as-you-go. The figures below are dated reference estimates, not guaranteed savings; check current provider rates before deciding.
DeepSeek V4.1 Flash
USD per 1 million tokens
| Plan / provider | Price / 1M tokens |
|---|---|
| ProSubscription · effective price | $0.0039 |
| MaxSubscription · effective price | $0.0043 |
| Pro MaxSubscription · effective price | $0.0034 |
| UltraSubscription · effective price | $0.0042 |
| Point packsPay as you go |
|
| DeepSeek APIOfficial · Off-peak / peak |
|
| Together AIFlash 1017 · Serverless · Precision not disclosed |
|
| Fireworks AIFlash 1017 · Serverless · Precision not disclosed |
|
| OpenRouterMorph · FP8 · Flash 1017 |
|
| SiliconFlowFP8 · Global · Flash |
|
| DeepInfraFP8 · Flash 1017 · Standard · 51% off |
|
| Novita AIFP8 · Flash 1017 · Serverless |
|
Cache read · Reusing input
A cache hit reuses computation for an identical input prefix, such as a repeated system prompt or conversation history. Only the matched input tokens receive the cache-read rate; the rest use the regular input rate. It does not reuse a finished answer or change the model’s numerical precision. Cache availability and rules vary by provider. About Cache →
FP8 · Numerical precision
FP8 means 8-bit floating point, used for parts of model inference to reduce memory use and potentially speed up computation. Compared with FP16/BF16, it can introduce numerical differences; the effect on answer quality depends on the model and implementation. A cache hit can happen on an FP8 route too: caching concerns input reuse, while FP8 concerns how numbers are represented. About FP8 →
Subscription effective price = monthly fee ÷ (7-day token allowance × 30 ÷ 7), shown per 1M tokens. Yearly plans use the annual fee ÷ 12. This assumes full weekly allowance use within the plan’s usage windows; using fewer tokens increases your effective price. Input, cached input and output share one allowance at 1×, so the same effective price applies to each.
Compare a workload, not just a token price.
Illustrative 30-day workloads with 90% input and 10% output. Each volume is shown without caching and with 80% of input cached. These are examples, not measured customer usage.
Annual plan: monthly equivalent; full year paid upfront. All amounts below are USD for the whole example workload.
| Total tokens / 30 days | Official API Off-peak / peak | Point-pack usage | Pro | Max | Pro Max | Ultra |
|---|---|---|---|---|---|---|
| 10M0% of input cached | $1.95 / $3.90 | $1.77 | $9.95 | $24.95 | $34.95 | $149.50 |
| 10M80% of input cached | $0.89 / $1.78 | $0.87 | $9.95 | $24.95 | $34.95 | $149.50 |
| 100M0% of input cached | $19.50 / $39.00 | $17.65 | $9.95 | $24.95 | $34.95 | $149.50 |
| 100M80% of input cached | $8.92 / $17.83 | $8.65 | $9.95 | $24.95 | $34.95 | $149.50 |
| 1,000M0% of input cached | $195.00 / $390.00 | $176.50 | $9.95 | $24.95 | $34.95 | $149.50 |
| 1,000M80% of input cached | $89.16 / $178.32 | $86.50 | $9.95 | $24.95 | $34.95 | $149.50 |
Subscription figures assume the workload fits every 5-hour and 7-day window and all other operating limits. A 30-day average cannot establish that fit. Reference weekly allowances × 30 ÷ 7 are used only for comparison, not as monthly entitlements. Point-pack amounts are usage charges before pack bonuses, not a checkout charge; purchased packs may leave credit unused. Taxes, payment fees, extra tools and overage are excluded.
API cost = uncached input × input rate + cached input × cache rate + output × output rate, using millions of tokens. Official columns assume all tokens use off-peak or all use peak rates; a mixed schedule falls between those amounts. Verify current official rates and hours ↗
What if I use only part of my allowance?
Effective subscription price in USD per 1M tokens at 10%, 30% or 100% of the normalized allowance. The subscription fee stays the same when you use less.
| Plan | 10% used | 30% used | 100% used |
|---|---|---|---|
| Pro | $0.0387 | $0.0129 | $0.0039 |
| Max | $0.0431 | $0.0144 | $0.0043 |
| Pro Max | $0.034 | $0.0113 | $0.0034 |
| Ultra | $0.0415 | $0.0138 | $0.0042 |
Unused quota is not a cash balance. Actual effective price is undefined when no tokens are used.
For low or irregular usage, pay-as-you-go may cost less. A subscription is worth considering when the recurring fee, verified features and usage windows fit your work. Read the plan selection guide →
Allowances remain reference values. Point-pack rates are deducted from API credit before pack bonuses. Official rates checked November 5, 2026. DeepSeek prices & peak hours ↗ · Plan limits ↗
Provider prices and precision labels checked November 5, 2026; click a provider for its source. OpenRouter uses the named Morph route; other providers show direct API prices. DeepInfra included a 51% promotion at the last check; its end date was not confirmed, and its cached price is rounded as published. SiliconFlow uses its global Flash listing. Token prices exclude credit-purchase fees and taxes. FP8 labels describe the provider’s published precision, not a measured quality comparison.