DeepSeek raised API prices on 16 August, and the same weights are cheaper on a provider that gives credits
V4-Flash output went from $0.28 to $0.66 off-peak and $1.32 at peak. Together AI serves the same MIT weights at $0.28 and runs a startup credit program.
DeepSeek raised its API prices at 16:00 UTC on 16 August and dropped the flat rate for peak and off-peak billing. A million output tokens from V4-Flash cost $0.28 before the change; they now cost $0.66 off-peak and $1.32 at peak. V4-Pro went from $0.87 to $1.98 off-peak and $3.96 at peak.
Peak hours are 01:00–04:00 and 06:00–10:00 UTC. For seven hours a day everything you send costs twice what it costs the rest of the time. If evaluation runs, bulk labelling or archive reprocessing are triggered by cron, look at when that cron fires: moving the job out of those two windows halves the line item and changes nothing else about it.
The second question is which invoice this lands on. No program in our catalogue pays for DeepSeek's own API, so the increase comes out of cash. But the model is not tied to that API — V4-Pro and V4-Flash are published on Hugging Face under the MIT licence, so anyone can serve them.
Together AI lists V4-Flash at $0.14 per million input tokens and $0.28 per million output, which is the rate DeepSeek itself charged until 16 August. Against the first-party API that is 2.4 times cheaper off-peak and 4.7 times cheaper at peak, and Together is one of the programs we track, so it can be paid with credits rather than cash.
Pro does not work out the same way. Together lists V4-Pro at $1.32 and $3.96 — exactly DeepSeek's new peak price, and worse than its $1.98 off-peak. For Pro the fix is the schedule rather than the provider. One detail also cuts the other way for both models: DeepSeek still bills cached input at $0.007 per million off-peak against Together's $0.03, so a workload dominated by cache hits keeps an advantage where it is.
What Together grants and who qualifies is on the Together AI card. Which credits cover open weights and which only pay for a provider's own models is in our guide to AI model credits.