DeepSeek V4 GA: pricing, архитектура CSA+HCA и gap vs GPT-5.6
peak-valley × SWE-bench 80.6% × API sunset 24 июля (2026)

Если в проде ещё крутится deepseek-chat или вы оцениваете DeepSeek V4 GA20 июля 2026 выкатывается full release: peak-valley pricing, SWE-bench Verified 80.6% (open-source SOTA), 1M token context window и hard sunset legacy API 24 июля. Разбор по исходнику без воды: timeline 24.04→20.07, V4-Pro/Flash stack (CSA+HCA+mHC+Muon), benchmark table vs GPT-5.6 / Claude Fable 5, полный price sheet (CNY/USD), 6-step API migration + code samples, pain points, cost hacks и FAQ. Updated: 2026-07-20

01

DeepSeek V4 GA: 6 pain points для integration teams

Три месяца preview → production GA. Performance bump реальный, но commercialization + breaking API changes = operational friction. Вот что чаще всего всплывает в инженерных чатах:

  1. 01

    Legacy countdown: deepseek-chat и deepseek-reasoner умирают 24 июля 15:59 UTC (23:59 Beijing) — немигрированный prod traffic = 404.

  2. 02

    Peak-valley complexity: первый time-of-use pricing — будни 09–12 и 14–18 (Beijing) x2 rate; 24/7 agent pipelines ломают cost forecasting.

  3. 03

    Closed-source всё ещё топ на hard benchmarks: Claude Fable 5 — 80.3% SWE-bench Pro; V4-Pro — 55.4%. Multi-file repo surgery = paid frontier.

  4. 04

    Self-host hardware floor: V4-Pro (1.6T/49B active) не влезет в laptop; V4-Flash local нужен 96GB+ UMA (см. antirez ds4 + Mac 96GB benchmark).

  5. 05

    Три inference modes: Non-think / Think High / Think Max — Think Max жрёт 384K+ context quota, cost + latency spike.

  6. 06

    От «почти free» к regulated platform: даже после –75% в июне peak-valley делает scheduling core competency, не просто price compare.

Timeline: preview → GA

ДатаСобытие
2026-04-24V4 preview + open weights (MIT): V4-Pro (1.6T), V4-Flash (284B)
2026-05Production-tuned V4-Flash/V4-Pro; API GA
2026-06V4-Pro permanent –75% (output $0.87/M tokens)
2026-06-29Email всем API users: GA mid-July + первое раскрытие peak-valley
2026-07-19Gray-test access; медиа: «full release завтра»
2026-07-20GA release (publish date)
2026-07-24deepseek-chat / deepseek-reasoner permanent offline

«Preview уже был сильным — GA = targeted bump в agent/math/code + formal peak-valley billing stack.»

02

V4-Pro vs V4-Flash: CSA + HCA под 1M context

DeepSeek V4 дропнул MLA из V2/V3 → CSA + HCA hybrid attention, mHC residuals, Muon optimizer. 1M context становится servable, а не marketing slide.

SpecV4-ProV4-Flash
Total params1.6 trillion (1.6T)284 billion (284B)
Active per token49 billion (49B)13 billion (13B)
Transformer layers6143
Context window1,000,000 tokens1,000,000 tokens
Max output384K tokens384K tokens
PrecisionFP4 (experts) + FP8 (rest)FP4 + FP8 mixed
Pretrain data33T+ tokens32T+ tokens
LicenseMITMIT

① Compressed Sparse Attention (CSA)

  • KV sequence softmax-gated pooling → compression
  • FP4 «lightning indexer» top-k: V4-Pro top-1024, V4-Flash top-512
  • Sliding window последних 128 tokens
  • При 1M context inference FLOPs = 27% от V3.2

② Heavily Compressed Attention (HCA)

  • Token compression 128× → dense global attention для long-range deps
  • V4-Flash: первые 2 слоя HCA, дальше CSA/HCA alternate
  • KV cache memory = 10% от V3.2 (V4-Flash: 7%)

③ mHC + Muon

  • mHC (Manifold-Constrained Hyper-Connections): 4-channel residual stream, doubly stochastic mix matrix (Birkhoff polytope) — stable signal через 61 layers
  • Muon optimizer: Newton-Schulz orthogonalized gradients вместо AdamW — faster convergence

Три inference modes + sampling params

ModeПоведениеUse case
Non-thinkБез CoT, fastest TTFTSimple Q&A, routing
Think HighExplicit reasoning (<redacted_thinking>)Medium complexity, code debug
Think MaxMax reasoning, нужен 384K+ contextHard math, long-chain agents

Official sampling (all modes): temperature=1.0, top_p=1.0

03

Benchmarks: SWE-bench 80.6% и 116× cost gap

V4-Pro бьёт open-weight records — closed frontier всё ещё ahead на hardest tasks. Решающая метрика: cost per task.

BenchmarkDeepSeek V4-ProClaude Fable 5GPT-5.6 UltraClaude Opus 4.8
SWE-bench Verified80.6% ★ open-source record96.0%Not published separately~69%
SWE-bench Pro55.4%80.3%78.1%69.2%
LiveCodeBench (Pass@1)93.5%88.1%87.4%83.2%
Codeforces Elo3,206
Terminal-Bench 2.183.9%88.0%85.1%82.7%

SWE-bench Verified = real GitHub bugfixes. V4-Pro 80.6% делит open-source #1 с Gemini 3.1 Pro. SWE-bench Pro жёстче — Claude Fable 5 лидирует 80.3%.

Cost-performance (Artificial Analysis)

  • Claude Fable 5: Strategy & Ops index — $3.48/task, score 50
  • DeepSeek V4-Pro: same bucket — $0.03/task, score 38
  • DeepSeek V4-Flash: все 6 index types < $0.04/task
  • Cost multiplier: Fable 5 = 116× дороже V4-Pro при ~12 points (~31%) score delta
info

Sources: DeepSeek docs, arXiv:2606.19348, HuggingFace, Artificial Analysis, LLMReference.

04

vs GPT-5.6 Sol / Claude Fable 5: decision matrix

DimensionDeepSeek V4-ProGPT-5.6 SolClaude Fable 5
Open source✅ MIT❌ Closed❌ Closed
Self-host✅ Yes
Context1M tokensNot disclosed1M tokens
Code capabilityVery strong (near Fable 5)Very strongStrongest
API output (off-peak)$0.87/M tokens~$15/M$50/M
API output (peak)$1.74/M tokens
Agent orchestrationStrong, multi-agentStrongStrongest
Data privacy✅ Private deploy
  • Tight budget / high QPS / self-host: V4-Pro или V4-Flash
  • Max code quality, cost secondary: Claude Fable 5 (SWE-bench Pro leader)
  • Hard algorithms + math: GPT-5.6 Sol / Ultra
  • Massive logs/docs on the cheap: V4-Flash cache hit $0.0028/M

Peak-valley price sheet (CNY + USD)

GA headline change: peak hours (Beijing weekdays 09:00–12:00 & 14:00–18:00) — rates x2.

ModelLine itemOff-peakPeak
V4-ProInput (cache hit)¥0.025 / $0.0035 / 1M tokens
Input (cache miss)¥3.00 / $0.435 / 1M tokens¥6.00 / $0.87
Output¥6.00 / $0.87 / 1M tokens¥12.00 / $1.74
V4-FlashInput (cache hit)¥0.02 / $0.0028 / 1M tokens
Input (cache miss)¥1.00 / $0.14 / 1M tokens¥2.00 / $0.28
Output¥2.00 / $0.28 / 1M tokens¥4.00 / $0.56

4 cost hacks

  1. 01

    Batch off-peak: doc processing, labeling, code review после 18:00 или до 09:00 (Beijing).

  2. 02

    Prompt cache: static system prompts first — V4-Flash cache hit $0.0028/M.

  3. 03

    Flash routing: simple intents → V4-Flash, hard reasoning → V4-Pro — save 60–80%.

  4. 04

    Billing emails: DeepSeek даёт 24h notice на price changes.

Даже в peak: V4-Pro output ($1.74/M) всё ещё 8.6× дешевле Claude Opus 4.8 ($15/M) и GPT-5.6 Sol (~$15/M).

05

API migration: 6 steps до deadline 24 июля

warning

Hard deadline: deepseek-chat и deepseek-reasoner отключаются 24 июля 2026, 15:59 UTC (23:59 Beijing).

Old modelNew modelNotes
deepseek-chatdeepseek-v4-flash (non-think)Fast, light tasks
deepseek-reasonerdeepseek-v4-flash (think) or deepseek-v4-proPro for stronger reasoning
  1. 01

    Repo-wide grep: deepseek-chat, deepseek-reasoner + env vars + CI configs.

  2. 02

    Mapping: light chat → deepseek-v4-flash; old reasoner → flash think or deepseek-v4-pro.

  3. 03

    OpenAI SDK: swap model only; think mode via extra_body (см. code).

  4. 04

    Anthropic SDK: base_url = https://api.deepseek.com, update model.

  5. 05

    Staging regression: прогнать agent/RAG до 24 июля, verify peak-valley bill.

  6. 06

    Prod canary: traffic split, rollback switch до deadline.

  7. 07

    Think Max: context window 384K+.

python
# ❌ Legacy (dead after July 24)
client.chat.completions.create(
    model="deepseek-chat",
    messages=[...]
)

# ✅ New
client.chat.completions.create(
    model="deepseek-v4-pro",          # or deepseek-v4-flash
    messages=[...],
    # Think mode:
    # extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)
python
# Anthropic SDK path
client = anthropic.Anthropic(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)
message = client.messages.create(
    model="deepseek-v4-pro",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Hello, DeepSeek V4!"}]
)

Hard numbers (EEAT)

  • KV cache @ 1M: V4-Pro 10%, V4-Flash 7% of V3.2
  • Inference FLOPs @ 1M: CSA → 27% of V3.2
  • June price cut: V4-Pro output $0.87/M (75% off), then peak-valley
  • Concurrency default: V4-Pro 500 parallel requests

DeepSeek V4 GA режет API bill, но local V4-Flash всё ещё требует 96GB UMA Mac; дешёвый Linux VPS не запустит xcodebuild/notarytool. Если вы крутите cloud DeepSeek API + iOS CI/CD / CLI agent long sessions — macOS heavy load на dedicated remote node стабильнее, чем bet на local iron. API pricing не чинит Keychain isolation и SSH session drops. Для prod с stable SSH, predictable bandwidthNodeMini Mac Mini cloud rental обычно лучший execution layer независимо от peak-valley swings. Specs: цены аренды; onboarding: help center; local flash: antirez ds4 + 96GB Mac benchmark.

Updated: 2026-07-20. Закройте API migration до 24 июля.

FAQ

FAQ

Архитектура та же; GA bump в agent/math/code + peak-valley. Legacy models sunset 24 июля. Budget: цены аренды.

Beijing weekdays 09:00–12:00 & 14:00–18:00 (x2). Batch после 18:00, prompt cache, V4-Flash routing.

deepseek-chatdeepseek-v4-flash (non-think); deepseek-reasoner → flash think or deepseek-v4-pro. Deadline 24 июля 23:59 Beijing.

SWE-bench Verified: V4-Pro 80.6% (open-source record), Fable 5 96%. Cost/self-host: V4-Pro $0.87/M ≈ 1/17 GPT-5.6 Sol. Remote env: help center.

Да. ds4 на 96GB UMA Mac (284B/13B active); V4-Pro (1.6T) → cloud vLLM/SGLang. Details: antirez ds4 Mac benchmark.

Simple Q&A: Non-think; code debug: Think High; hard math/long agents: Think Max (384K+). Sampling: temperature=1.0, top_p=1.0.