DEV Community

Owen
Owen

Posted on Originally published at ofox.ai

DeepSeek API Price Increase: Up to 12x, Peak Hours (2026)

DeepSeek API Price Increase: Up to 12x, Peak Hours (2026)

DeepSeek implemented significant API pricing changes effective August 16, 2026 at 16:00 UTC, introducing a peak/off-peak billing model replacing previous flat rates. The most dramatic increase affected V4 Pro cache hits, which jumped 12.1x at peak ($0.003625 to $0.044 per million tokens).

Key Changes

Effective Date: 2026-08-16, 16:00 UTC (2026-08-17, 00:00 Beijing time)

Peak Hours (7 hours daily): 01:00–04:00 and 06:00–10:00 UTC

Billing Model: Off-peak rates are 50% of peak rates across all tiers

Pricing Comparison Table

Tier Old Flat New Off-peak New Peak
V4 Flash cache hit $0.0028 $0.007 $0.014
V4 Flash cache miss $0.14 $0.22 $0.44
V4 Flash output $0.28 $0.66 $1.32
V4 Pro cache hit $0.003625 $0.022 $0.044
V4 Pro cache miss $0.435 $0.66 $1.32
V4 Pro output $0.87 $1.98 $3.96

Price Increase Multipliers

Tier Off-peak vs Old Peak vs Old
V4 Flash cache hit 2.5x 5.0x
V4 Flash cache miss 1.6x 3.1x
V4 Flash output 2.4x 4.7x
V4 Pro cache hit 6.1x 12.1x
V4 Pro cache miss 1.5x 3.0x
V4 Pro output 2.3x 4.6x

Real-World Impact: Sample Monthly Bill

For a typical agent workload (300M input tokens at 90% cache hit rate, 20M output):

Scenario Cache Hits Cache Misses Output Total
Old flat rate $0.76 $4.20 $5.60 $10.56
New, all off-peak $1.89 $6.60 $13.20 $21.69
New, all peak $3.78 $13.20 $26.40 $43.38

Peak Hours by Time Zone (Aug 2026)

Location Peak Window 1 Peak Window 2 Workday Impact
Beijing/Singapore (UTC+8) 09:00–12:00 14:00–18:00 7 of 9 hours
Tokyo/Seoul (UTC+9) 10:00–13:00 15:00–19:00 6 of 9 hours
India (UTC+5:30) 06:30–09:30 11:30–15:30 4.5 of 9 hours
Berlin/Paris (UTC+2, CEST) 03:00–06:00 08:00–12:00 3 of 9 hours
London (UTC+1, BST) 02:00–05:00 07:00–11:00 2 of 9 hours
New York (UTC-4, EDT) 21:00–00:00 prev 02:00–06:00 0 of 9 hours
San Francisco (UTC-7, PDT) 18:00–21:00 prev 23:00–03:00 0 of 9 hours

Third-Party Hosts Comparison

V4 Flash (on OpenRouter, 2026-08-17)

Host Input Output Cache Read
DeepInfra $0.08 $0.18 $0.016
DigitalOcean $0.08 $0.252 $0.0252
GMICloud $0.084 $0.168 $0.0168
DeepSeek $0.44 $1.32 $0.014

Finding: 26 of 28 third-party listings undercut DeepSeek's peak input rate.

Break-even analysis: At 94% cache hit rate off-peak, DeepInfra becomes cheaper; at 99.4% peak. Previously, this threshold was 77%.

V4 Pro

DeepSeek's off-peak rate ($0.66 input) undercuts all nine available third-party listings.

Historical Context

DeepSeek has adjusted pricing three times in eighteen months:

Date Change Direction
2025-02-26 Off-peak discounts introduced (16:30–00:30 UTC) Down
2025-09-05 New pricing, off-peak discounts ended Up
2026-08-16 Peak/off-peak billing (off-peak = 50% of peak) Up

This represents the first instance where time-of-day pricing increased rates rather than offered discounts.

Why Cache Hits Increased Most

The previous cache hit rate was an outlier: V4 Flash hits at $0.0028 were 50x cheaper than a miss. That disproportionate pricing supported the cost model for agent loops with 90%+ cache hit rates. The new structure reflects market realignment, where caching remains valuable but no longer dominates billing.

Caching Still Viable?

Yes. A cache hit remains approximately 31x cheaper than a miss at peak ($0.014 vs $0.44 for Flash). However, this advantage narrowed from the previous 50x ratio. Output tokens now represent a larger portion of total costs.

Key Considerations

Timestamp Ambiguity: DeepSeek has not clarified whether requests billing during peak windows are determined by start time or completion time. Requests crossing 04:00 or 10:00 UTC boundaries represent an open question.

Off-Peak as Partial Relief: Off-peak rates represent a smaller increase, not a true discount. Even at off-peak, V4 Flash output costs $0.66 (2.4x the old flat $0.28).

No Grandfather Clause: Existing balances top up at new rates; no legacy pricing applies.

Gateway Strategy: Using an OpenAI-compatible API aggregator allows switching between providers via model string changes alone, avoiding lock-in to any single vendor's pricing schedule.

Frequently Asked Questions

Does the increase differ between V4 Flash and V4 Pro?

Yes substantially. Cache hit increases ranged from 5x (Flash) to 12.1x (Pro) at peak.

How does billing handle requests crossing peak boundaries?

Unknown. DeepSeek's documentation omits this; measure actual invoices for verification.

Are third-party hosts now cheaper?

For V4 Flash: nearly uniformly yes (26 of 28 endpoints). For V4 Pro: no — DeepSeek undercuts all current alternatives even at off-peak rates.

Is prompt caching still worthwhile?

Absolutely, though with diminished leverage. The economics shifted toward output token optimization.

Did previous balances retain old rates?

No documented grandfather provision exists. All consumption applies current pricing.


Originally published on ofox.ai/blog.

Top comments (0)