DeepSeek V4 Pro Goes GA, and LLM APIs Get Their First Peak-Hour Pricing

DeepSeek quietly pushed V4 Pro 0813 to general availability this week, and buried in the same announcement is a change that matters more to working engineers than the model itself: starting August 17, DeepSeek API pricing splits into peak and off-peak rates, tied to the wall clock in UTC. This is the first time a frontier-adjacent LLM provider has shipped time-of-day pricing, and if it works, expect every other provider watching their GPU utilization curves to copy it within a quarter.

The model release is genuinely strong on its own. V4 Pro 0813 is a 1.6 trillion parameter mixture-of-experts model with roughly 49 billion active parameters per token, pretrained on more than 32 trillion tokens, and it ships with a 1,048,576 token context window and up to 384,000 output tokens. A hybrid attention design keeps inference cost from exploding at that context length. But it is the pricing model sitting next to it that changes how you should architect anything that calls this API in volume.

Three things worth understanding#

Peak hours are fixed windows, not a load-based surge. DeepSeek defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC, every day, regardless of actual demand. Everything outside those windows is off-peak and billed at half the peak rate. For V4 Pro, output tokens move from a flat $0.87 per million to $3.96 per million at peak and $1.98 per million off-peak. That is a predictable, schedulable cost curve rather than the unpredictable surge pricing you would get from a spot-market model, which makes it something you can actually plan a batch pipeline around.

The agentic benchmark gains are the real headline for anyone building agents. Vendor-reported numbers show DeepSWE jumping from 12.8 to 62.7 and Terminal-Bench 2.1 climbing from 72.1 to 87.9 versus the preview build. Those are the benchmarks that track multi-step tool use and code execution, exactly the workloads that tend to run long and rack up token spend, which makes the timing of the pricing change look deliberate rather than coincidental.

No independent verification exists yet. As of this week, no third party had reproduced DeepSeek’s benchmark claims. Every number circulating is vendor-reported. That is not unusual for a same-week GA launch, but it means the intelligence-index comparisons you will see in roundup posts should be read as marketing until someone outside DeepSeek runs the eval suite.

Try it: peak vs off-peak cost estimator#

If you run agentic workloads against this kind of pricing, the question that matters is simple: how much does shifting a job’s schedule actually save you? Enter a token count below and see the peak, off-peak, and old flat-rate cost side by side, using V4 Pro’s published output-token pricing.

Enter a token count and click Estimate.

Where this leaves API consumers#

If DeepSeek’s off-peak windows hold up in practice, the move rewards exactly the workloads that do not need a synchronous response: batch evaluation runs, nightly agent sweeps, offline document processing. Anything user-facing and latency-sensitive is stuck paying the peak rate regardless of your engineering effort, which quietly shifts the incentive toward async-first agent design. My take is that this is a preview of where the whole industry is headed as inference capacity becomes the real bottleneck rather than model quality. The practical move today is to tag your own workloads by whether they are latency-sensitive or not, and start routing the batchable half of your agent traffic toward off-peak windows before every provider adds this same lever and the free lunch of flat-rate batching disappears.