DeepSeek V4 Flash GA: бенчмарки у Opus 4.8, цена — доля процента

~16 мин чтения · MACCOME · Обновлено: 5 августа 2026

Аудитория: AI-разработчики, agent engineers, tech leads с API budget и routing SLA. 31 июля 2026 DeepSeek выпустил V4-Flash-0731 official: 284B / 13B active без изменений — прирост только post-training; agent scores обогнали V4-Pro preview; input $0.14/M (cache miss), output $0.28/M — порядок 1/36–1/179 от Claude Opus 4.8. Получите: timeline, pricing matrix шести игроков, разбор CSA+HCA+mHC+Muon, Harness mechanism, Artificial Analysis cross-check, controversy boundaries, 6-step runbook, FAQ. Структура: blind spots → timeline → pricing → architecture → Harness → competition → controversy → runbook → wrap. GA V4 и peak pricing: V4 GA migration guide; Kimi K3: Kimi K3 open weight release.

bolt

TL;DR — 30 секунд

  • Не новая архитектура: 284B / 13B active идентичны апрельскому preview — performance delta = post-training only, не parameter inflation.
  • Price floor: input $0.14/M (miss) / $0.0028/M (hit), output $0.28/M — ~1/36–1/179 vs Opus 4.8 в зависимости от cache hit rate.
  • Harness debut: in-house agent runtime vs Claude Code; не опубликован; все vendor agent scores — minimal mode, max tier, top_p=0.95, temp=1.0.
  • V4-Pro official — нет: на 5.08. только Flash-0731 (API-only); Pro official + Harness — «скоро», без даты.
  • Benchmark caveat: Terminal Bench 2.0 82.7 — Harness-dependent; не экстраполируйте на Claude Code / Cursor без independent rerun.

Шесть blind spots после GA: что ломает routing и FinOps

После 31.07 команды упираются в шесть механических ограничений — каждое бьёт по bill или agent reliability:

  1. Harness lock-in на бенчмарках: Terminal Bench 2.0 82.7 — unpublished Harness minimal mode; vendor changelog прямо: «scores highly harness-sensitive». Не равно generic Claude Code throughput.
  2. V4-Pro official — TBD: 1.6T / 49B active preview живёт, official GA без даты; медиа «10–20 августа» — unsourced; changelog: «скоро».
  3. API-only split: 0731 official только через API; consumer app/web на старом build — multi-endpoint teams получают inconsistent behavior.
  4. Cache miss tax: overseas dev feedback — низкий prompt cache hit rate; agent loops с fixed system prompt могут платить x50 vs list price при miss-heavy traffic.
  5. Peak surcharge pending: DeepSeek анонсировал 2x pricing Beijing 9–12 и 14–18; effective date на 5.08. не опубликована — batch scheduler должен держать buffer.
  6. Funding noise: ~$7.4B raise / ~$48.7B valuation — press «по слухам», без regulatory filing; не factor в production routing.

Timeline: preview апреля → Flash-0731 official

ДатаСобытие
2026-04-24V4 preview + open weights: V4-Pro (1.6T / 49B active), V4-Flash (284B / 13B active), 1M context, MIT
2026-07-24Deprecated deepseek-chat и deepseek-reasoner отключены; весь traffic на V4 naming
2026-07-27Moonshot публикует Kimi K3 (2.8T) weights на Hugging Face — давление на DeepSeek pricing
2026-07-31V4-Flash-0731 official в API beta; weights sync; changelog первый раз называет Harness agent framework
2026-08-02Alibaba Qwen3.8-Max API GA — ещё один open-weight competitor в CN stack
2026-08-05 (cutoff)V4-Pro official — нет; Harness — не public; app/web не на 0731

Pricing matrix: Flash-0731 vs шесть CN/open players

Vendor-published на 5.08.2026. GLM-5.2 pricing — не verified в рамках этой статьи.

МодельСтатусTotal / activeContextInput (miss/hit, $/M)Output ($/M)License
V4-Flash-0731Official GA284B / 13B1M$0.14 / $0.0028$0.28MIT
V4-ProPreview1.6T / 49B1M$0.435 / $0.003625$0.87MIT
Kimi K3Open weights2.8T / ~104B (community est.)~1.05M$3.00 / $0.30$15.00Modified MIT
GLM-5.2Open source~744B / ~40B1Mне verifiedне verifiedMIT
Qwen3.8-MaxAPI GA2.4T / 95B1M$2.00 / ~$0.17–0.25$6.00open-source promised
GPT-5.6 SolClosedне раскрыто———proprietary
Claude Fable 5Closedне раскрыто1M—~$50proprietary

Architecture frozen, post-training moved: CSA + HCA + mHC + Muon

Ключевой mechanism insight: V4-Flash-0731 — тот же parameter graph, что апрельский preview. 284B total / 13B active per token обогнал 1.6T V4-Pro preview на agent benchmarks — сигнал H2 2026: post-training ROI может превысить raw scale-up.

По tech report DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence:

  • Hybrid attention (CSA + HCA): marketed as DSA sparse attention; снижает compute + memory bandwidth на long context. При 1M tokens V4-Pro inference FLOPs = 27% от V3.2, KV Cache = 10% (vendor-reported; independent third-party rerun pending).
  • Manifold-constrained hyper-connections (mHC): стабилизирует gradient/signal flow в deep residual stack — меньше degradation на глубоких слоях.
  • Muon optimizer: замена AdamW; faster convergence, меньше training instability на MoE routing.

Harness: agent execution layer DeepSeek vs Claude Code

Changelog 31.07 первый раз называет DeepSeek Harness — in-house runtime: file I/O, tool calls, shell, end-to-end engineering tasks. Positioning: direct competitor Claude Code. До этого DeepSeek models жили в Claude Code, OpenCode, сторонних harness.

Все vendor agent scores (Terminal Bench 2.0, Toolathlon) — Harness minimal mode, max tier, top_p=0.95, temperature=1.0. Vendor disclaimer: «benchmark results highly sensitive to harness selection» — score != raw model capability delta.

Artificial Analysis cross-check: intelligence не top, cost per task — floor

МодельLabIntelligence Index (third party)Avg cost per task (third party)
V4-Flash-0731DeepSeek50$0.03
Kimi K3Moonshot AI57$0.86
GLM-5.2Zhipu / Z.ai~+1 pt vs Flashне verified
GPT-5.6 SolOpenAI+9 pts vs Flash$1.86
Claude Fable 5Anthropic+9 pts vs Flash$3.15

Source: Artificial Analysis independent eval, cited via finance media. Methodology != DeepSeek Harness agent scores — не merge datasets.

Mechanism vs market: V4-Flash не лидер Intelligence Index (trails K3, GLM-5.2), но per-task cost ~1/29 K3, ~1/62 GPT-5.6 Sol, ~1/105 Claude Fable 5. DeepSeek продаёт good-enough intelligence at extreme $/token — объясняет 7 недель #1 на OpenRouter call volume.

Controversy boundaries: что vendor-reported, что operational

  • Agent scores Harness-bound: Terminal Bench 2.0 82.7 vs V4-Pro preview 67.9 — unpublished framework; directional only до Claude Code / Cursor rerun.
  • Operational friction: per 21st Century Business Herald / overseas dev reports — low cache hit rate, occasional safety-classifier timeout на production traffic.
  • V4-Pro / Harness GA dates: CN media window 10–20 августа — unsourced; official: coming soon.
  • Funding figures: ~$7.4B / ~$48.7B valuation — press без regulatory confirmation; не input для infra capex.

Market mechanism: price floor и «kill line»

Пока V4-Pro official slip, CN community flip nickname founder Liang Wenfeng: «Liang Baikai» (ждать зря) → praise после Flash-0731. Kill line concept: DeepSeek ставит floor — конкурентам нужен и performance win, и lower $/M, иначе token share уходит. Контекст: OpenAI GPT-5.6 Luna −80% в июле (OpenAI pricing breakdown).

Launch day 31.07: NVIDIA, Broadcom, AMD — flat. Конtrast с DeepSeek-R1 shock января 2025: рынок привык к efficiency narrative без chip panic.

6-step runbook: evaluate → route → production

  1. Model pointer: API deepseek-v4-flash → 0731 official auto. Retire deepseek-chat — migration: V4 GA migration guide.
  2. Staging A/B: real prompt subset 0731 vs preview; focus tool-call JSON schema и long-context KV stability.
  3. Tier routing Flash / Pro: batch + router → Flash ($0.28/M output); hard reasoning → V4-Pro preview до official GA.
  4. Cache hit monitoring: fixed system prompt agent loops — enable Prompt Cache; miss-heavy = bill >> list price.
  5. Peak scheduling buffer: non-real-time batch вне Beijing 9–12, 14–18 после активации 2x surcharge.
  6. Local vs cloud: 128GB Mac — Flash q2 через ds4; production 24/7 agents — dedicated node: ds4 local vs cloud rental decision.
python
from openai import OpenAI

client = OpenAI(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)

# deepseek-v4-flash → 0731 official
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Разбери этот agent log..."}],
)

Три hard numbers для memo

  • 82.7 vs 67.9 — Terminal Bench 2.0: V4-Flash-0731 vendor score > V4-Pro preview (Harness minimal mode; independent rerun pending)
  • 27% / 10% — при 1M context V4-Pro inference FLOPs и KV Cache vs V3.2 (DeepSeek tech report, vendor-reported)
  • 1/29 ~ 1/105 — V4-Flash per-task cost vs Kimi K3 / GPT-5.6 Sol / Claude Fable 5 (Artificial Analysis third-party)

Wrap: good-enough intelligence, extreme $/token — production agents нужен stable host

V4-Flash-0731 не бьёт Claude Fable 5 / GPT-5.6 Sol на raw intelligence ceiling. Value proposition: MIT-licensed open weights + API at 1/36–1/179 closed flagship pricing для bulk agents и high-frequency calls. Для budget-sensitive teams с private deploy или heavy API volume — strongest value pick августа 2026.

Production DeepSeek V4 agent stack (OpenClaw Gateway, future Harness hooks, multi-model routing) упирается в три structural limits на laptop:

  • Lid-close sleep: Wi-Fi drop + sleep kill long agent sessions; spent tokens не refundable.
  • Cache + peak cron: local scheduler не держит off-peak batch при sleep policies.
  • No 24/7 router: Flash / Pro / Fable 5 tiering требует always-on Gateway; notebook — плохой dispatch node.

Для stable 24/7 agent automation MACCOME Mac cloud hosts — real macOS, SSH handoff, isolated env. Тарифы: цены аренды Mac mini.

Sources: DeepSeek API docs + changelog, tech report, Hugging Face model cards, Artificial Analysis (via finance media), 21st Century Business Herald, V2EX. Cutoff 5 августа 2026 — verify pricing и V4-Pro / Harness status перед production cutover.

FAQ

Главное отличие DeepSeek V4 от V3.2?

Нативный 1M-token context. При этом окне inference FLOPs V4-Pro — 27% от V3.2, KV Cache — 10% (vendor-reported). V4 серия — agent-focused tuning под Claude Code, OpenCode и аналоги.

V4 Flash или V4 Pro для daily use?

Chat, batch, agent pipelines — V4-Flash-0731 official (дешевле, agent scores выше V4-Pro preview). Сложный reasoning — V4-Pro preview до official GA.

V4 Pro official уже доступен?

На 5 августа 2026 — нет. GA только V4-Flash-0731 (API-only); app/web не на 0731. V4-Pro official + Harness — coming soon; слухи «10–20 августа» unverified.

Можно ли доверять бенчмаркам DeepSeek?

Tier trust: SWE-bench Verified и third-party baseline — выше. Terminal Bench 2.0 — Harness minimal mode; vendor warns harness-sensitivity. Ждите independent Claude Code / Cursor rerun.

Как гонять DeepSeek V4 agents 24/7 в production?

OpenClaw Gateway или multi-model routing на dedicated Mac cloud host. См. цены аренды Mac mini MACCOME для нод и конфигурации.