After three months of preview testing, DeepSeek V4 GA (general availability) went live on July 20, 2026. Compared with the April preview, the GA release upgrades Agent performance, math reasoning, and code generation — and introduces peak-valley pricing for the first time, signaling DeepSeek's shift from near-free access toward disciplined commercial operations.

One-line summary: The preview was already strong; GA adds targeted Agent, math, and code upgrades plus a formal commercial billing structure. The underlying MoE architecture is unchanged — GA focuses on stability, optimization, and pricing maturity.

DeepSeek V4 delivers three architectural breakthroughs that cut KV-cache memory to 10% of V3.2 at 1M-token context (V4-Flash as low as 7%), making ultra-long context economically viable.

Hybrid attention: CSA + HCA

DeepSeek V4 replaces V2/V3 MLA with a hybrid of two new attention mechanisms:

Manifold-constrained hyper-connections (mHC) and Muon optimizer

mHC upgrades standard residual connections with a 4-channel residual stream and mixing matrices constrained to the Birkhoff polytope (doubly stochastic), keeping signals stable across 61 layers. The Muon optimizer (replacing AdamW) applies Newton-Schulz orthogonalization to gradients for faster convergence and more stable training.

SWE-bench Verified tests real GitHub bug fixes. 80.6% is the current open-source high, tied with Gemini 3.1 Pro. SWE-bench Pro is stricter — Claude Fable 5's 80.3% still leads there.

Scenario picks: Tight budget / high volume / private deploy → V4-Pro or V4-Flash; maximum code quality, cost secondary → Claude Fable 5; hard algorithms and math → GPT-5.6 Sol / Ultra; massive log or document ingestion → V4-Flash (cached input from $0.0028/M).

The biggest GA billing change is peak-valley differentiated pricing, similar to time-of-use electricity rates. Peak hours (Beijing time): weekdays 09:00–12:00 and 14:00–18:00 — rates double.

Four cost-saving tips: (1) Schedule non-real-time jobs off-peak (after 18:00 or before 09:00 Beijing time); (2) Build Prompt Cache for repeated system prefixes; (3) Route lightweight work to Flash, escalate hard reasoning to Pro; (4) Watch DeepSeek email alerts 24 hours before billing changes. Even at peak, V4-Pro output at $1.74/M remains 8.6x cheaper than Claude Opus 4.8 ($15/M).

Six migration steps:

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

DeepSeek V4 GA is one of 2026's most important open-model milestones. Its value is not beating Claude Fable 5 or GPT-5.6 on every chart today — it is delivering undisputed open-source peak performance at roughly 1/10 to 1/100 of closed flagship cost.

For teams that cannot send data offshore, indie developers on tight budgets, Agent engineers needing 1M context, and AI products chasing cost efficiency, DeepSeek V4 Pro remains the most balanced choice. Peak-valley billing adds planning overhead but stays highly competitive — a maturity signal, not a retreat from value.

Lay out the alternatives: running DeepSeek Agent load tests on a personal Mac hits sleep and network interruptions on long-context jobs; closed APIs alone cannot match V4's MIT license plus $0.87/M off-peak output on cost and data sovereignty; missing the July 24 migration cutoff causes hard production outages. For iOS CI/CD, 24/7 DeepSeek V4 Agent automation, and local ds4 inference validation, KVMNODE dedicated Mac Mini M4 cloud rental is usually the better host: Apple Silicon unified memory, open sudo, flexible daily/weekly/monthly terms. See the pricing page, help center, or order directly.

Data as of: July 20, 2026 · Sources: DeepSeek official docs, arXiv:2606.19348, Hugging Face, Artificial Analysis