DeepSeek Is Raising API Prices on August 16: The Exact Numbers
Quick answer. DeepSeek is raising API prices from 16:00 UTC on August 16, 2026, replacing flat rates with peak/off-peak billing. V4-Pro output goes from $0.87 to $1.98 off-peak and $3.96 at peak — up to 4.55x. Cache-hit input rises up to 12x, which hits agentic workloads hardest. Peak hours are 01:00–04:00 and 06:00–10:00 UTC.
DeepSeek published a pricing change alongside its V4-Pro general-availability release, and it has been almost entirely overlooked in the coverage. From 16:00 UTC on August 16, 2026, DeepSeek moves to peak and off-peak billing — and both tiers are more expensive than today's flat rate.
If you run DeepSeek in production, you have a short window to model the impact and adjust. Here are the exact numbers.
What are the current prices?
Until 16:00 UTC on August 16, DeepSeek charges a single flat rate per model:
| Model | Cache hit | Cache miss (input) | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 |
All figures are per million tokens. Both models carry a 1M-token context window and a 384K maximum output.
What are the new prices?
| Model | Tier | Cache hit | Cache miss | Output |
|---|---|---|---|---|
| deepseek-v4-flash | Off-peak | $0.007 | $0.22 | $0.66 |
| Peak | $0.014 | $0.44 | $1.32 | |
| deepseek-v4-pro | Off-peak | $0.022 | $0.66 | $1.98 |
| Peak | $0.044 | $1.32 | $3.96 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC. That is seven hours a day; the other seventeen are off-peak.
How much more will you pay?
This is the table that matters. Every multiplier below is against today's flat rate.
| Model / token type | Off-peak | Peak |
|---|---|---|
| Pro — cache hit | 6.1x | 12.1x |
| Pro — cache miss | 1.5x | 3.0x |
| Pro — output | 2.3x | 4.6x |
| Flash — cache hit | 2.5x | 5.0x |
| Flash — cache miss | 1.6x | 3.1x |
| Flash — output | 2.4x | 4.7x |
There is no cheaper option than today. Off-peak is not a discount off the current price — it is half of a newly raised peak. Even running exclusively in off-peak hours, V4-Pro output costs 2.3 times what it does now.
Why the cache-hit increase is the real story
The headline numbers understate the impact for the workload most developers care about.
Cache-hit input on V4-Pro goes from $0.003625 to $0.044 at peak — a 12.1x increase. Off-peak is still 6.1x.
Cache hits are where agentic coding lives. Every turn of an agent loop re-sends the accumulated conversation, and that resend is what the cache is for. A long-running agent session is overwhelmingly cache-hit tokens, which is exactly why DeepSeek's near-zero cache rate made it so attractive for that workload.
Put plainly: if you chose DeepSeek because agent loops were almost free, that specific advantage shrinks by roughly an order of magnitude. Standard input tokens, by contrast, only rise 1.5x off-peak — the increase is deliberately concentrated on the tokens that heavy users generate most of.
A worked example
Take a modest agentic workload on V4-Pro: 50M cache-hit input, 10M cache-miss input, and 10M output tokens per month.
| Scenario | Cache hit | Cache miss | Output | Monthly total |
|---|---|---|---|---|
| Today | $0.18 | $4.35 | $8.70 | $13.23 |
| All off-peak | $1.10 | $6.60 | $19.80 | $27.50 |
| All peak | $2.20 | $13.20 | $39.60 | $55.00 |
Roughly 2.1x more if you schedule everything off-peak, and 4.2x if you do not. Scale those numbers to your own volume — the ratios hold.
Even after the rise, DeepSeek remains cheap in absolute terms against Claude Opus 5 at $5/$25 or Grok 4.6 at $2/$6. The point is not that DeepSeek becomes expensive. It is that the gap narrows sharply, and any cost model you built before August 16 is now wrong.
What should you do before August 16?
- Re-run your cost model. Anything built on the flat rate is obsolete in three days. Pay particular attention to cache-hit volume — that line will move most.
- Check your traffic against the peak windows. Peak is 01:00–04:00 and 06:00–10:00 UTC. If you are in Asia-Pacific, a meaningful share of your working day falls inside that band; if you are in the Americas, most of your traffic is naturally off-peak.
- Shift batch work into off-peak hours. Anything asynchronous — evaluations, bulk summarisation, offline refactors, test generation — should be scheduled outside the peak windows. That is a straightforward 50% saving on work that does not care when it runs.
- Re-examine Flash versus Pro. Flash stays roughly 3x cheaper than Pro at every tier. If Pro was marginal for your task before, the wider absolute gap may now justify Flash.
- Re-run your competitor comparison. Off-peak V4-Pro at $0.66/$1.98 still undercuts almost everything. At peak, $1.32/$3.96 sits close to Muse Spark 1.2 at $1.25/$4.25 — and the comparison is no longer obvious.
Is the price rise justified?
Worth noting for context: this lands with the V4-Pro general-availability release, which brought real capability upgrades — agent improvements, low/high/max reasoning-effort control on both models, and native OpenAI Responses API support with one-click Codex integration.
DeepSeek has also run time-based pricing before. Through 2025 it offered off-peak discounts on V3 and R1. What has changed is the direction: this time off-peak is half of an increased peak rather than a reduction from a standard rate. The mechanism is familiar; the economics are not.
Read charitably, it is a company that priced aggressively to win adoption and is now moving toward sustainable rates while giving schedulable workloads a genuine lever. Read plainly, it is a price rise of between 1.5x and 12x depending on the token type, announced three days ahead.
Both readings are true. The actionable part is the same either way: model your own numbers before Sunday.
FAQ
When does DeepSeek's new pricing take effect?
16:00 UTC on August 16, 2026.
What are DeepSeek's peak hours?
01:00–04:00 and 06:00–10:00 UTC — seven hours a day. All other hours bill at the off-peak rate.
How much is DeepSeek V4-Pro after the change?
Off-peak: $0.022 cache hit, $0.66 cache miss, $1.98 output per million tokens. Peak: $0.044, $1.32 and $3.96. Today's flat rate is $0.003625, $0.435 and $0.87.
How much is DeepSeek V4-Flash after the change?
Off-peak: $0.007, $0.22 and $0.66 per million tokens. Peak: $0.014, $0.44 and $1.32. Today's flat rate is $0.0028, $0.14 and $0.28.
Is off-peak cheaper than the current price?
No. Off-peak is half of the new peak rate, not a discount off today's flat rate. Every tier is more expensive than current pricing — V4-Pro output is 2.3x more even off-peak.
Which token type increases most?
Cache-hit input on V4-Pro, rising 6.1x off-peak and 12.1x at peak. That is the dominant token type in agentic loops, so agent workloads are affected far more than the headline rates suggest.
Is DeepSeek still cheap after the increase?
Yes, in absolute terms. Off-peak V4-Pro at $0.66/$1.98 still undercuts Claude Opus 5 ($5/$25) and Grok 4.6 ($2/$6). But at peak it approaches Muse Spark 1.2's $1.25/$4.25, so the comparison is much closer than it was.
How can I reduce the impact?
Schedule asynchronous work outside 01:00–04:00 and 06:00–10:00 UTC for a 50% saving, re-evaluate whether V4-Flash is sufficient, and reduce redundant context in agent loops now that cache hits are no longer near-free.