DeepSeek V4 GA (июль 2026): цены, бенчмарки и сравнение с GPT-5.6

~18 мин чтения · MACCOME · Обновлено: 20 июля 2026

Для кого: инженеры inference, agent-разработчики и те, кто копает механику MoE + attention, а не пресс-релизы. 20 июля 2026 — GA DeepSeek V4: те же веса апрельского preview, но production-tuned agent/math/code, peak-valley pricing и hard sunset legacy endpoints 24 июля. Внутри: timeline, CSA/HCA/mHC/Muon, benchmark-таблицы vs GPT-5.6 Sol и Claude Fable 5, pricing grid, migration mapping, 6 decision modules, 6-step runbook. Локальный inference: ds4 на Mac 128GB.

bolt

TL;DR — механика в 30 секунд

  • GA 20.07: V4-Pro 1.6T / 49B active ; V4-Flash 284B / 13B active — оба 1M context, MIT.
  • KV trick: CSA+HCA → 27% FLOPs и 10% KV vs V3.2 на 1M token — это не marketing, это arithmetic.
  • SWE-bench Verified 80.6% — open-weight record ; LiveCodeBench 93.5% > Fable 5 / GPT-5.6.
  • Off-peak output: Pro $0.87/M, Flash $0.28/M ; peak ×2 (Pekin weekdays 09–12, 14–18).
  • Migration: deepseek-chat → deepseek-v4-flash до 24.07 15:59 UTC.

Шесть модулей: где ломается production routing

  1. Модуль 1 — Legacy alias: deepseek-chat мапится на Flash, но alias умрет 24.07 — grep по репо обязателен.
  2. Модуль 2 — Peak multiplier: cron в Pekin business hours = 2× bill ; batch inference сдвинуть на off-peak.
  3. Модуль 3 — MoE tiering: Flash router (13B active/token) для bulk ; Pro (49B) для think-max chains — иначе переплата за activated params.
  4. Модуль 4 — Benchmark harness mismatch: SWE-bench Pro 80.3% у Fable 5 vs 55.4% у V4-Pro — но $0.03/task vs $3.48 ; смотрите metric, который матчит SLA.
  5. Модуль 5 — Context economics: без CSA/HCA 1M window = OOM ; с ними KV 10% — можно реально держать full repo в одном forward pass.
  6. Модуль 6 — Self-host path: MIT weights + ds4/vLLM ; API direct = data egress в CN infra — для sensitive codebases только private gateway.

Timeline: preview → GA → sunset

ДатаСобытие
24.04.2026Preview + MIT weights: V4-Pro (1.6T), V4-Flash (284B)
Май 2026Production-tuned API endpoints
Июнь 2026V4-Pro output -75% permanent ($0.87/M off-peak)
29.06.2026Email: GA mid-July + peak-valley pricing
19.07.2026Gray GA access для selected devs
20.07.2026GA live
24.07.2026deepseek-chat / deepseek-reasoner EOL (15:59 UTC)

Архитектура: spec sheet и механизмы

SpecV4-ProV4-Flash
Total params1.6T284B
Active / token49B (~3%)13B (~4.6%)
Layers6143
Context1M1M
Max output384K384K
PrecisionFP4 experts + FP8 restFP4 + FP8
Pretrain33T+ tokens32T+ tokens
LicenseMITMIT

CSA (Compressed Sparse Attention)

Pipeline: KV sequence → 4× softmax-gated pool → FP4 lightning indexer picks top-1024 blocks (Pro) / top-512 (Flash) → + 128-token sliding window для recency bias. На 1M context: inference FLOPs = 27% от V3.2 full attention baseline.

HCA (Heavily Compressed Attention)

128× compression → dense global attention over compressed cache. CSA/HCA alternate per layer (не parallel) — local sparse + global dense без quadratic KV blowup.

mHC (Manifold-Constrained Hyper-Connections)

4-channel residual stream + doubly stochastic mix matrix (Birkhoff polytope) — gradient stability через 61 layer без vanilla x + f(x) degradation.

Muon optimizer

Newton-Schulz orthogonalization на gradient steps — faster convergence vs AdamW на 33T+ pretrain.

Три inference mode

ModeMechanismWhen
Non-thinkNo CoT tokens emittedRouter, classification, low-latency
Think HighExplicit <redacted_thinking> blocksDebug, medium reasoning
Think MaxMax budget, needs 384K+ ctxMath, multi-hop agents

Sampling (official): temperature=1.0, top_p=1.0 — all modes.

Benchmarks: numbers, not vibes

BenchmarkV4-ProFable 5GPT-5.6 UltraOpus 4.8
SWE-bench Verified80.6%96.0%—~69%
SWE-bench Pro55.4%80.3%78.1%69.2%
LiveCodeBench Pass@193.5%88.1%87.4%83.2%
Codeforces Elo3206———
Terminal-Bench 2.183.9%88.0%85.1%82.7%

Cost/task (Artificial Analysis): Fable 5 — 50 pts, $3.48 ; V4-Pro — 38 pts, $0.03 → 116× cheaper, 24% score gap.

Peak-valley pricing grid

ModelItemOff-peak USD/MPeak USD/M
V4-ProInput cache hit0.00352×
V4-ProInput cache miss0.4350.87
V4-ProOutput0.871.74
V4-FlashInput cache hit0.00282×
V4-FlashInput cache miss0.140.28
V4-FlashOutput0.280.56

Peak window: Pekin weekdays 09:00–12:00, 14:00–18:00. Savings playbook: off-peak cron, prompt cache (Flash hit $0.0028/M input), Flash→Pro escalation, email alerts 24h ahead.

V4-Pro vs GPT-5.6 Sol vs Fable 5

AxisV4-ProGPT-5.6 SolFable 5
Open weightsMIT, self-hostClosedClosed
1M contextYes (CSA/HCA)UnconfirmedYes
Output off-peak$0.87/M~$15/M~$50/M
Terminal agents83.9%85.1%88.0%
LiveCodeBench93.5%87.4%88.1%

API migration (deadline 24.07.2026)

LegacyGA equivalentNotes
deepseek-chatdeepseek-v4-flash non-thinkDrop-in chat
deepseek-reasonerdeepseek-v4-flash think or deepseek-v4-proPro = stronger chain
python
from openai import OpenAI

client = OpenAI(api_key="KEY", base_url="https://api.deepseek.com/v1")

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Fix this bug..."}],
    extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)

6-step runbook

  1. grep legacy: deepseek-chat|deepseek-reasoner across services.
  2. Define routing graph: Flash bulk → Pro think → GPT-5.6 terminal-only paths.
  3. Staging cutover with GA model IDs ; measure cache hit % and p95 latency.
  4. Shift batch jobs off Pekin peak windows.
  5. Self-host eval: ds4 on 128GB Mac cloud for Flash ; vLLM for Pro if GPU budget exists.
  6. Prod deploy before 24.07 ; keep rollback branch until smoke tests pass.

Verdict + MACCOME bridge

GA — это не новая архитектура, а production seal на MoE stack с реально работающим 1M context (10% KV) и open-weight SWE record 80.6%. GPT-5.6 Sol и Fable 5 держат agent ceiling ; V4-Pro держит $/token floor.

7×24 agent gateway на ноутбуке = sleep + unified memory contention. MACCOME Mac cloud: native macOS, SSH, ds4/OpenClaw isolation — private inference без laptop scheduler. Тарифы: аренда Mac mini.

Sources: DeepSeek docs, arXiv:2606.19348, HuggingFace, Artificial Analysis. 20.07.2026.

FAQ

Когда отключат deepseek-chat?

24 июля 2026, 15:59 UTC. Замените на deepseek-v4-flash или deepseek-v4-pro.

Off-peak цена V4-Pro output?

$0.87/M token. Peak ×2. TCO gateway node vs API-only: тарифы MACCOME.

Как CSA снижает KV cache?

4× pool + FP4 indexer (top-1024/512) + 128-token window. На 1M ctx KV = 10% от V3.2 (Flash 7%).

V4-Pro или GPT-5.6 Sol?

V4-Pro: cost + self-host + LiveCodeBench 93.5%. GPT-5.6: Terminal-Bench 85.1%, Ultra subagents. Hybrid routing — default для production.

Flash vs Pro routing?

Flash: 13B active, cache hit $0.0028/M input — bulk/router. Pro: 49B active, SWE 80.6% — think chains. Tiered = 60–80% cost cut.

7×24 V4 agents?

OpenClaw Gateway или ds4-server на dedicated Mac cloud — no sleep. MACCOME аренда.