먼저 결론
- 2026-08-16 이후 deepseek-v4-flash는 UTC 피크/오프피크. 오프피크는 정확히 절반.
- 오프피크 miss in $0.22/M, out $0.66/M; 피크 2배. hit 오프피크 in $0.007/M.
- 8/16 이전 $0.14/$0.28 통일가 폐지.
- 분수령은 출력 token 비중 × UTC 창. thinking 기본 ON, 동가.
- 1M 컨텍스트 + OpenAI 호환
{API_BASE}. 피크 스케줄과 cache가 레버.
모델 ID가 분수령이 아님. UTC 피크/오프피크와 출력 비중이 분수령.
0. 먼저 결론
2026-08-22 DeepSeek official pricing (effective 2026-08-16 16:00 UTC) splits deepseek-v4-flash (checkpoint DeepSeek-V4-Flash-0731) into peak/off-peak × cache-hit/miss × input/output. Base https://api.deepseek.com, OpenAI-compatible; 1M context; thinking on by default at same rates.
Off-peak Flash: cache-miss in $0.22/M, out $0.66/M; peak doubles to $0.44 / $1.32. Cache-hit off-peak in $0.007/M. Peak UTC 01:00–04:00 and 06:00–10:00. Legacy flat $0.14/$0.28 dead. Gemini compare: Gemini API 2026; stack: AI Agent stack.
1. 왜 단가가 하나가 아닌가
Teams budgeted pre-Aug 16 flat rates, launched into UTC peak, and output at $1.32/M blew budgets. Others quoted off-peak input while agent loops emit ≥3× output; thinking billed the same. No cache on 80k system prompts is expensive at peak miss $0.44/M; cache-hit off-peak gap ≈31×. August 2026 adds a time axis—not Gemini Batch, not Claude seats. Claude: Claude Code 2026.
- Peak is whole-table ×2; sessions can span both windows.
- Cache-hit vs miss ≈31× off-peak—skipping cache is repeat-input tax.
- 1M context is a cap, not free quota.
- Thinking defaults on; same price.
- MCP retries at peak rebill output. Host: MCP placement.
2. 네 가지 과금 차원
2.1 Peak vs off-peak (UTC)
Peak 01:00–04:00 and 06:00–10:00 UTC; off-peak half price. Cron batch to off-peak.
2.2 Cache-hit vs miss
Stable prefixes hit at Flash off-peak $0.007/M.
2.3 Input vs output
Off-peak out $0.66 is 3× miss in $0.22; peak out $1.32 is 94× hit in $0.014. Output share is the main dial.
2.4 Flash vs Pro
Pro off-peak out $1.98 ≈3× Flash; router only.
3. 같은 헤더 비교
| 층/방안 | 진입 | 실행 | 컨텍스트 | 비용 | 권한 경계 |
|---|---|---|---|---|---|
| Off-peak Flash | https://api.deepseek.com | 1M ctx | cache tiers | in $0.007–0.22, out $0.66 | Keys backend |
| Peak Flash | UTC peak stamp | same | same | in $0.014–0.44, out $1.32 | Time-sliced |
| Legacy flat | pre–8/16 | none | none | in $0.14, out $0.28 | obsolete |
| Off-peak Pro | deepseek-v4-pro | stronger | cache | in $0.022–0.66, out $1.98 | hard cases |
| Cache | stable prefix | cut miss | system+Schema | hit −97% | change=remiss |
| Host | Cloud Mac/VPS | MCP, sign | repos | peak retry waste | always-on |
비대칭 결론: 피크→오프피크 50% 절약, 출력 비중 70%→40%도 비슷 이상. 둘 다 필요.
3.1 공식 치트시트
| Model | Window | hit in | miss in | out | Note |
|---|---|---|---|---|---|
| deepseek-v4-flash | Off-peak | $0.007 | $0.22 | $0.66 | 0731 |
| deepseek-v4-flash | Peak | $0.014 | $0.44 | $1.32 | UTC 01–04, 06–10 |
| deepseek-v4-pro | Off-peak | $0.022 | $0.66 | $1.98 | peak ×2 |
| (legacy) | Flat | $0.0028 | $0.14 | $0.28 | dead |
4. 시나리오
| 층/방안 | 진입 | 실행 | 컨텍스트 | 비용 | 권한 경계 |
|---|---|---|---|---|---|
| Demo | Flash off-peak | short | <8k | <$5/mo | test keys |
| Daily product | Flash+cache | cap tools | short ctx | out 3× in | UTC monitor |
| Code agent | Flash; Pro route | thinking same | repo summary | out+retry | MCP on |
| Night batch | off-peak cron | chunk 1M | JSONL | out $0.66 | queue policy |
| Long RAG | high cache hit | static cached | short query | miss controlled | version prefix |
| iOS agent | Cloud Mac+MCP | infer only | certs on Mac | avoid peak retry | keychain stays |
5. 추천 조합
A Personal: Flash off-peak, short prompts, not legacy $0.14/$0.28 B Product: Flash+cache, batch cron off-peak, Pro router only C Multi-model: DeepSeek + Gemini/Claude split (stack article) D Apple: DeepSeek infer; MCP/sign on Cloud Mac
6. 오해
- 1: Pre–Aug 16 flat rates—obsolete.
- 2: Input-only quotes; output $0.66 off / $1.32 peak.
- 3: Thinking-off discount—none; same card.
- 4: No cache on large system prompts.
- 5: Ignoring UTC 06:00–10:00 peak band.
- 6: Unbounded retries at peak.
- 7: Default Pro when Flash suffices.
7. 7단계
- Open official page; confirm 2026-08-16 16:00 UTC and peak windows.
- Run prod prompt off-peak and peak; log cache hit/miss.
- Cache system prompt and tool Schema; target >80% hit.
- Cap tool steps and daily USD; queue peak alerts to off-peak.
- Cron batch/evals to UTC off-peak.
- Move MCP off closing laptops. MCP + Cloud Mac.
- Internal rate card: off-peak/peak columns; diff invoices by UTC hour.
8. FAQ
DeepSeek V4-Flash 현재 가격?
Off-peak miss in $0.22/M, out $0.66/M; peak doubles. Hit off-peak in $0.007/M.
피크/오프피크 정의?
UTC 01:00–04:00 and 06:00–10:00 peak (2×). Billing uses API request UTC timestamp.
thinking과 non-thinking 동가?
Same price; thinking on by default.
Gemini와 비교?
No single sticker compare. High cache + off-peak batch favors DeepSeek input; peak output agents need $1.32/M. Gemini 2026.
model id?
deepseek-v4-flash (0731). Pro: deepseek-v4-pro. Base https://api.deepseek.com.
왜 클라우드 Mac?
DeepSeek sells inference only. Peak retries rebill output at $1.32/M. Cloud Mac cuts wasted output tokens.
9. 요약
2026년 8월 이후 DeepSeek V4-Flash API 가격은 UTC 피크/오프피크 × cache × 입출력 비율. 출력 70%면 스케줄링 alone ~35% 절약. 출력·cache·피크 재시도를 함께.