DeepSeek V4 Full Release (July 2026):
Pricing, Benchmarks & How It Stacks Up Against GPT-5.6

Real pain points before you switch:

• Legacy endpoints expire July 24: deepseek-chat and deepseek-reasoner will be permanently disabled. Unmigrated production code breaks immediately.
• Peak pricing adds scheduling overhead: Rates double on weekdays 09:00–12:00 and 14:00–18:00 Beijing time. 24/7 agent pipelines need rescheduling.
• GA performance differs from preview: Agent, math, and code capabilities received targeted improvements. April benchmarks are no longer reliable.
• Pro vs Flash routing matters: 1.6T and 284B specs are very different. Wrong routing wastes budget or sacrifices quality.

The GA release is not a new architecture — it is the April preview graduating into production with stability improvements and a formal commercial pricing structure.

DeepSeek replaced Multi-head Latent Attention (MLA) from V2/V3 with a two-track hybrid system:

• Compressed Sparse Attention (CSA): Compresses KV entries 4x via softmax-gated pooling, then uses an FP4 lightning indexer for top-k selection (top-1024 on Pro, top-512 on Flash), plus a 128-token sliding window. At 1M tokens, inference needs only 27% of V3.2 FLOPs.
• Heavily Compressed Attention (HCA): 128x compression followed by dense global attention. Captures long-range patterns CSA might miss. CSA and HCA alternate across layers.
• Manifold-Constrained Hyper-Connections (mHC): A 4-channel residual stream governed by a doubly stochastic matrix keeps gradients stable through 61 deep layers.
• Muon Optimizer: Newton-Schulz orthogonalization conditions gradient steps better than AdamW, enabling faster convergence.

Net result: KV cache at 1M tokens drops to 10% of V3.2 memory (7% on V4-Flash).

Official recommended sampling: temperature=1.0, top_p=1.0 for all modes.

Per Artificial Analysis on Strategy & Ops tasks: Claude Fable 5 scores 50 at $3.48 per task; DeepSeek V4-Pro scores 38 at $0.03 per task. That is 116x cheaper for a 24% performance gap.

Pick V4-Pro or V4-Flash for budget, high-volume, or self-hosted workloads. Pick Claude Fable 5 for the hardest multi-file repo work. Pick GPT-5.6 for terminal-heavy agentic workflows and M365/Azure integration.

The most significant GA change: time-based pricing. Rates double during Beijing weekday peak hours: 09:00–12:00 and 14:00–18:00.

Four practical cost cuts:

• Schedule batch jobs off-peak: Run bulk summarization, code review, and data pipelines after 18:00 or before 09:00 Beijing time.
• Maximize cache hit rate: Static system prompts first. V4-Flash cache hits cost just $0.0028/M.
• Route simple queries to V4-Flash: Intent classification and FAQ responses on Flash; escalate complex reasoning to Pro.
• Monitor billing emails: DeepSeek sends 24-hour advance notice on pricing changes.

Even at peak, V4-Pro output at $1.74/M is 8.6x cheaper than Claude Opus 4.8 ($15/M) and GPT-5.6 Sol (~$15/M).

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

# Old (stops working July 24)
client.chat.completions.create(
    model="deepseek-chat",
    messages=[...]
)

# New
client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[...],
    # extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)

V4 supports both OpenAI ChatCompletions and Anthropic Messages API formats. Base URL stays the same; only update the model name.

Primary official sources — re-check links after any upstream update:

https://api-docs.deepseek.com

https://arxiv.org/abs/2606.19348

https://huggingface.co/deepseek-ai

Six-step migration checklist:

1. Search your codebase: Find all references to deepseek-chat and deepseek-reasoner in code, CI configs, and env vars.
2. Map replacements: Lightweight chat to deepseek-v4-flash (Non-think); reasoning to Flash thinking mode or deepseek-v4-pro.
3. Update SDK calls: OpenAI-compatible clients only need a model name change. Anthropic SDK keeps base_url=https://api.deepseek.com.
4. Configure thinking mode: Former reasoner users enable thinking via extra_body. Think Max requires 384K+ context.
5. Test in staging: Complete end-to-end validation before July 24. Verify peak pricing impact on your budget.
6. Cron off-peak batch jobs: Schedule document processing and code review after 18:00 or on weekends. Combine with Prompt Cache for maximum savings.

Citable technical data:

• SWE-bench Verified 80.6%: Best open-weight score, tied with Gemini 3.1 Pro (source: DeepSeek official docs, Jul 2026).
• KV cache at 10%: 1M-token scenarios use only 10% of V3.2 memory; V4-Flash as low as 7% (source: arXiv:2606.19348).
• 116x cost advantage: Artificial Analysis Strategy & Ops: V4-Pro $0.03/task vs Fable 5 $3.48/task (source: Artificial Analysis, Jul 2026).

Bottom line: DeepSeek V4 GA is not about beating Claude Fable 5 or GPT-5.6 on every benchmark — it is about delivering the strongest open-weight performance at 1/10 to 1/100 the cost of closed-source frontier models.

When wiring DeepSeek V4 API into a local OpenClaw Gateway or Cursor agent stack, a sleeping Mac interrupts long SWE-bench-style runs, and peak-hour 24/7 pipelines are hard to schedule on a laptop. Pure API calls need no local GPU, but hybrid setups (local gateway + cloud inference + always-on orchestration) often lack a stable host for agent loops. For production environments running AI coding agents, OpenClaw multi-model routing, or 24/7 agent cycles, CALMVPS bare-metal Mac Mini rental is usually the better choice: dedicated Apple Silicon, Metal-native acceleration, flexible monthly billing, ~120-second delivery. See the pricing page.