DeepSeek V4 GA Release

Peak-valley pricing · 1M context · 80.6% SWE-bench · Migrate before July 24

DeepSeek V4 GA release pricing benchmarks architecture July 2026

After three months of preview access, DeepSeek V4 officially hit general availability on July 20, 2026. If you are asking is DeepSeek V4 better than GPT-5.6 or still calling deepseek-chat in production, this developer-focused guide covers the full timeline, V4-Pro/Flash architecture, benchmark tables, peak-valley pricing, an honest comparison with Claude Fable 5, and a six-step migration runbook before the July 24 deadline. GA brings agent/math/code upgrades plus time-based pricing — SWE-bench Verified hits 80.6% (open-weight record) at $0.87/M output off-peak. Data as of July 20, 2026.

01

What Changed in the GA Release vs. the April Preview

The preview dropped April 24, 2026. Today's GA is not a new architecture — it graduates the preview into production with stability, optimization, and commercial pricing discipline.

DateMilestone
Apr 24, 2026Preview + MIT open source: V4-Pro (1.6T) and V4-Flash (284B)
May 2026Production-tuned versions; API generally available
Jun 2026V4-Pro output permanently cut 75% to $0.87/M tokens
Jun 29, 2026Email to all API users: GA mid-July + first peak-valley pricing disclosure
Jul 19, 2026Gray-release testers; media reports imminent full launch
Jul 20, 2026GA live globally
Jul 24, 2026deepseek-chat and deepseek-reasoner permanently disabled (15:59 UTC)

Five pain points developers face right now:

  1. 01

    Legacy endpoints dying in 4 days: deepseek-chat and deepseek-reasoner stop working July 24.

  2. 02

    Peak pricing complexity: Weekday business-hour rates double — 24/7 pipelines need scheduling.

  3. 03

    Closed-model costs: Claude Fable 5 at ~$50/M output and GPT-5.6 Sol at ~$15/M drain budgets at scale.

  4. 04

    1M context on paper only: Most models choke on KV cache memory at long context.

  5. 05

    Data sovereignty: Closed APIs cannot be self-hosted for compliance-sensitive workloads.

02

Architecture: How DeepSeek Makes 1M Tokens Practical

SpecV4-ProV4-Flash
Total parameters1.6 trillion284 billion
Active per token49B (3%)13B (4.6%)
Layers6143
Context window1,000,000 tokens1,000,000 tokens
Max output384K384K
PrecisionFP4 (MoE experts) + FP8Same
Training data33T+ tokens32T+ tokens
LicenseMITMIT

CSA + HCA Hybrid Attention

V4 replaced MLA from V2/V3 with a two-track system. Compressed Sparse Attention (CSA) compresses KV 4x via softmax-gated pooling, uses an FP4 lightning indexer (top-1024 Pro / top-512 Flash), plus a 128-token sliding window. Heavily Compressed Attention (HCA) compresses 128x then runs dense global attention. Layers alternate CSA and HCA.

At 1M tokens: V4-Pro uses 27% of V3.2 inference FLOPs; KV cache drops to 10% of V3.2 memory (Flash: 7%).

mHC and Muon Optimizer

Manifold-Constrained Hyper-Connections (mHC) replace standard residuals with a 4-channel stream governed by a doubly stochastic matrix — stable gradients through 61 layers. Muon (not AdamW) uses Newton-Schulz orthogonalization for faster, stabler training.

Three Reasoning Modes

ModeUse case
Non-thinkFast routing, classification, simple Q&A
Think HighDebugging, medium complexity
Think MaxComplex math, multi-step agents (384K+ context)

Recommended sampling: temperature=1.0, top_p=1.0 for all modes.

03

Benchmark Numbers: Where V4 Wins and Where It Does Not

BenchmarkV4-ProClaude Fable 5GPT-5.6 UltraOpus 4.8
SWE-bench Verified80.6% open record96.0%N/A~69%
SWE-bench Pro55.4%80.3%78.1%69.2%
LiveCodeBench93.5%88.1%87.4%83.2%
Codeforces Elo3,206
Terminal-Bench 2.183.9%88.0%85.1%82.7%

Artificial Analysis Strategy & Ops index: Fable 5 scores 50 at $3.48/task; V4-Pro scores 38 at $0.03/task116x cheaper for a 24% performance gap. V4-Flash stays under $0.04/task across all six index categories.

04

DeepSeek V4 vs GPT-5.6 vs Claude Fable 5: Honest Comparison

DimensionV4-ProGPT-5.6 SolClaude Fable 5
Open / self-hostableYes (MIT)NoNo
1M contextYesNot confirmedYes
Best coding (hardest repos)Second tierSecond tierLeads SWE-bench Pro
Algorithms / mathLiveCodeBench leaderStrong
Output (off-peak)$0.87/M~$15/M~$50/M
Output (peak)$1.74/M
Data sovereigntySelf-host possibleNoNo

Peak-Valley Pricing Table

Peak hours (Beijing Time): Weekdays 09:00–12:00 and 14:00–18:00 — all rates double.

ModelItemOff-PeakPeak
V4-ProInput (cache hit)$0.0035/M2x
Input (cache miss)$0.435/M$0.87/M
Output$0.87/M$1.74/M
V4-FlashInput (cache hit)$0.0028/M2x
Input (cache miss)$0.14/M$0.28/M
Output$0.28/M$0.56/M

Even at peak, V4-Pro output is 8.6x cheaper than Claude Opus 4.8 ($15/M) and GPT-5.6 Sol (~$15/M).

Cost tips: Schedule batch jobs off-peak; maximize prompt cache hits; route simple queries to V4-Flash; monitor DeepSeek billing-change emails (24h notice).

05

Migration Runbook: Six Steps Before July 24

Deadline: July 24, 2026, 15:59 UTC. Legacy names deepseek-chat and deepseek-reasoner will be permanently disabled.

Old nameNew equivalent
deepseek-chatdeepseek-v4-flash (non-thinking)
deepseek-reasonerdeepseek-v4-flash (thinking) or deepseek-v4-pro
  1. 01

    Search codebase: Grep for deepseek-chat and deepseek-reasoner including env vars and CI configs.

  2. 02

    Pick target model: Flash for chat/routing; Pro for heavy reasoning.

  3. 03

    Update OpenAI SDK: Change model only; base_url stays https://api.deepseek.com.

  4. 04

    Enable thinking mode: Use extra_body with thinking config to replace reasoner behavior.

  5. 05

    Test in staging: Validate core flows; Think Max needs 384K+ context window.

  6. 06

    Ship before deadline: Anthropic SDK users swap model name only — same base URL.

python
response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Solve step by step..."}],
    extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)

Hard Numbers Worth Citing

  • SWE-bench Verified: 80.6% — best open-weight score
  • KV cache: 10% of V3.2 memory at 1M tokens
  • Cost ratio: $0.03 vs $3.48 per task vs Fable 5
  • Peak output: $1.74/M — still 8.6x below closed frontier APIs
  • Migration deadline: July 24, 2026 15:59 UTC

The GA release does not need to beat Claude Fable 5 on every benchmark. It delivers the strongest open-weight coding model at 1/10 to 1/100 the cost of closed frontier APIs.

Self-hosting V4 or running long-context agents still needs reliable hardware — laptops cannot stay up 24/7 and VMs add performance and compliance risk. For production iOS CI/CD and AI agent automation, VpsMesh Mac Mini cloud rental is usually the better fit: bare-metal Apple Silicon pairs well with local inference stacks like ds4 + V4 Flash on Mac.

Sources: DeepSeek official docs, arXiv:2606.19348, HuggingFace, LLMReference, Artificial Analysis, MangoMind Blog, TechNode.

FAQ

Frequently Asked Questions

Weekday Beijing hours 09:00–12:00 and 14:00–18:00 double all rates. V4-Pro off-peak output is $0.87/M; peak is $1.74/M. See our Mac Mini rental pricing for complementary dev environment costs.

Before July 24, replace deepseek-chat with deepseek-v4-flash and deepseek-reasoner with Flash thinking mode or deepseek-v4-pro. Base URL unchanged.

Pro is the 1.6T flagship (49B active); Flash is 284B (13B active). Both support 1M context. Flash is cheaper ($0.28/M output) for routing; Pro for complex agents.

Fable 5 and GPT-5.6 lead on hardest benchmarks. V4-Pro is MIT-licensed, self-hostable, and among the cheapest frontier AI APIs in 2026 at $0.87/M off-peak output.

V4-Flash can run via engines like ds4 on high-end Macs (96GB+ unified memory). Full Pro needs cluster hardware. Setup details in our help center.

Agent, math, and code performance upgrades; peak-valley pricing; global GA availability; legacy model names retire July 24. Underlying MoE architecture unchanged.