{"id":4337,"date":"2026-08-16T03:56:44","date_gmt":"2026-08-16T03:56:44","guid":{"rendered":"https:\/\/tucumandevelopers.com\/index.php\/2026\/08\/16\/deepseek-now-prices-tokens-like-electricity-50-off-peak-discount-and-a-spring-boot-pattern-to-profit-from-it\/"},"modified":"2026-08-16T03:56:44","modified_gmt":"2026-08-16T03:56:44","slug":"deepseek-now-prices-tokens-like-electricity-50-off-peak-discount-and-a-spring-boot-pattern-to-profit-from-it","status":"publish","type":"post","link":"https:\/\/tucumandevelopers.com\/index.php\/2026\/08\/16\/deepseek-now-prices-tokens-like-electricity-50-off-peak-discount-and-a-spring-boot-pattern-to-profit-from-it\/","title":{"rendered":"DeepSeek Now Prices Tokens Like Electricity: 50% Off-Peak Discount and a Spring Boot Pattern to Profit From It"},"content":{"rendered":"
\n
<\/div>\n

Then define the off-peak window exactly as DeepSeek defines it, in UTC: <\/p>\n

\n
@Component<\/span> public<\/span> class<\/span> OffPeakWindow<\/span> {<\/span> private<\/span> static<\/span> final<\/span> ZoneId<\/span> UTC<\/span> =<\/span> ZoneId<\/span>.<\/span>of<\/span>(<\/span>\"UTC\"<\/span>);<\/span> public<\/span> boolean<\/span> isOffPeak<\/span>(<\/span>ZonedDateTime<\/span> now<\/span>)<\/span> {<\/span> int<\/span> hour<\/span> =<\/span> now<\/span>.<\/span>withZoneSameInstant<\/span>(<\/span>UTC<\/span>).<\/span>getHour<\/span>();<\/span> \/\/ Peak: 01:00-04:00 and 06:00-10:00 UTC. Everything else is off-peak.<\/span> return<\/span> !((<\/span>hour<\/span> >=<\/span> 1<\/span> &&<\/span> hour<\/span> <<\/span> 4<\/span>)<\/span> ||<\/span> (<\/span>hour<\/span> >=<\/span> 6<\/span> &&<\/span> hour<\/span> <<\/span> 10<\/span>));<\/span> }<\/span> }<\/span> <\/code><\/pre>\n
\n<\/p><\/div>\n<\/p><\/div>\n
Now the batch job itself. The cron runs in UTC, so the schedule is stable regardless of where the server lives. 22:00 UTC is 04:00 in Dhaka, deep inside off-peak: <\/p>\n
\n
@Component<\/span> public<\/span> class<\/span> NightlyEmbeddingJob<\/span> {<\/span> private<\/span> static<\/span> final<\/span> Logger<\/span> log<\/span> =<\/span> LoggerFactory<\/span>.<\/span>getLogger<\/span>(<\/span>NightlyEmbeddingJob<\/span>.<\/span>class<\/span>);<\/span> private<\/span> final<\/span> OffPeakWindow<\/span> offPeakWindow<\/span>;<\/span> private<\/span> final<\/span> ChatClient<\/span> chatClient<\/span>;<\/span> public<\/span> NightlyEmbeddingJob<\/span>(<\/span>OffPeakWindow<\/span> offPeakWindow<\/span>,<\/span> ChatClient<\/span> chatClient<\/span>)<\/span> {<\/span> this<\/span>.<\/span>offPeakWindow<\/span> =<\/span> offPeakWindow<\/span>;<\/span> this<\/span>.<\/span>chatClient<\/span> =<\/span> chatClient<\/span>;<\/span> }<\/span> @Scheduled<\/span>(<\/span>cron<\/span> =<\/span> \"${ai.batch.cron}\"<\/span>,<\/span> zone<\/span> =<\/span> \"UTC\"<\/span>)<\/span> public<\/span> void<\/span> runBatch<\/span>()<\/span> {<\/span> if<\/span> (!<\/span>offPeakWindow<\/span>.<\/span>isOffPeak<\/span>(<\/span>ZonedDateTime<\/span>.<\/span>now<\/span>()))<\/span> {<\/span> log<\/span>.<\/span>warn<\/span>(<\/span>\"Skipping batch run: peak hours in UTC\"<\/span>);<\/span> return<\/span>;<\/span> }<\/span> \/\/ Heavy work: embeddings, summarization, eval runs.<\/span> }<\/span> }<\/span> <\/code><\/pre>\n
\n<\/p><\/div>\n<\/p><\/div>\n
\n
ai.batch.cron<\/span>=<\/span>0 0 22 * * *<\/span> <\/code><\/pre>\n
\n<\/p><\/div>\n<\/p><\/div>\n
The window check is a safety net, not the primary control. The cron is the primary control, and the check exists for the day someone changes the cron to 08:00 UTC and forgets what that costs. Defensive, cheap, and it turns a pricing policy into code.<\/p>\n
The second piece is an estimator, because a cost problem you cannot see is a cost problem you will not fix. Model the official table as data: <\/p>\n
\n
public<\/span> record<\/span> TokenPrice<\/span>(<\/span>double<\/span> inputPerM<\/span>,<\/span> double<\/span> outputPerM<\/span>,<\/span> double<\/span> cacheHitPerM<\/span>)<\/span> {}<\/span> public<\/span> final<\/span> class<\/span> DeepSeekRates<\/span> {<\/span> public<\/span> static<\/span> final<\/span> TokenPrice<\/span> FLASH_OFF_PEAK<\/span> =<\/span> new<\/span> TokenPrice<\/span>(<\/span>0.22<\/span>,<\/span> 0.66<\/span>,<\/span> 0.007<\/span>);<\/span> public<\/span> static<\/span> final<\/span> TokenPrice<\/span> FLASH_PEAK<\/span> =<\/span> new<\/span> TokenPrice<\/span>(<\/span>0.44<\/span>,<\/span> 1.32<\/span>,<\/span> 0.014<\/span>);<\/span> public<\/span> static<\/span> final<\/span> TokenPrice<\/span> PRO_OFF_PEAK<\/span> =<\/span> new<\/span> TokenPrice<\/span>(<\/span>0.66<\/span>,<\/span> 1.98<\/span>,<\/span> 0.022<\/span>);<\/span> public<\/span> static<\/span> final<\/span> TokenPrice<\/span> PRO_PEAK<\/span> =<\/span> new<\/span> TokenPrice<\/span>(<\/span>1.32<\/span>,<\/span> 3.96<\/span>,<\/span> 0.044<\/span>);<\/span> private<\/span> DeepSeekRates<\/span>()<\/span> {<\/span> }<\/span> }<\/span> <\/code><\/pre>\n
\n<\/p><\/div>\n<\/p><\/div>\n
Then a component that picks the right tier by hour and logs an estimate after every call: <\/p>\n
\n
@Component<\/span> public<\/span> class<\/span> CostTracker<\/span> {<\/span> private<\/span> final<\/span> OffPeakWindow<\/span> offPeakWindow<\/span>;<\/span> public<\/span> CostTracker<\/span>(<\/span>OffPeakWindow<\/span> offPeakWindow<\/span>)<\/span> {<\/span> this<\/span>.<\/span>offPeakWindow<\/span> =<\/span> offPeakWindow<\/span>;<\/span> }<\/span> public<\/span> void<\/span> track<\/span>(<\/span>String<\/span> model<\/span>,<\/span> TokenPrice<\/span> peak<\/span>,<\/span> TokenPrice<\/span> offPeak<\/span>,<\/span> long<\/span> cacheHitTokens<\/span>,<\/span> long<\/span> inputTokens<\/span>,<\/span> long<\/span> outputTokens<\/span>)<\/span> {<\/span> boolean<\/span> offPeak<\/span> =<\/span> offPeakWindow<\/span>.<\/span>isOffPeak<\/span>(<\/span>ZonedDateTime<\/span>.<\/span>now<\/span>());<\/span> TokenPrice<\/span> price<\/span> =<\/span> offPeak<\/span> ?<\/span> offPeak<\/span> :<\/span> peak<\/span>;<\/span> double<\/span> cost<\/span> =<\/span> (<\/span>cacheHitTokens<\/span> \/<\/span> 1_000_000.0<\/span>)<\/span> *<\/span> price<\/span>.<\/span>cacheHitPerM<\/span>()<\/span> +<\/span> (<\/span>inputTokens<\/span> \/<\/span> 1_000_000.0<\/span>)<\/span> *<\/span> price<\/span>.<\/span>inputPerM<\/span>()<\/span> +<\/span> (<\/span>outputTokens<\/span> \/<\/span> 1_000_000.0<\/span>)<\/span> *<\/span> price<\/span>.<\/span>outputPerM<\/span>();<\/span> log<\/span>.<\/span>info<\/span>(<\/span>\"{} {} cost estimate: ${} (tokens: {} cache hit \/ {} input \/ {} output)\"<\/span>,<\/span> model<\/span>,<\/span> offPeak<\/span> ?<\/span> \"off-peak\"<\/span> :<\/span> \"peak\"<\/span>,<\/span> cost<\/span>,<\/span> cacheHitTokens<\/span>,<\/span> inputTokens<\/span>,<\/span> outputTokens<\/span>);<\/span> }<\/span> }<\/span> <\/code><\/pre>\n
\n<\/p><\/div>\n<\/p><\/div>\n
Wire track(...)<\/code> wherever you call the model, and after a week you have a distribution: what your workload costs at each hour. That distribution is the thing to optimize, because the off-peak discount only helps the workloads you actually move.<\/p>\n
  <\/a> The checklist before you run a bill through it <\/h2>\n