DeepSeek V4 Full GA Is Live: Top Performance, Price Disruptor

Introduction

After nearly three months of waiting, DeepSeek V4 GA (full release) is finally here.

On July 20, 2026, the highly anticipated Chinese LLM began limited gray testing as DeepSeek V4 Flash and DeepSeek V4 Pro. Versus preview, the GA build not only lifts overall performance — it also introduces industry-disrupting peak-valley billing for the first time.

Meanwhile, the legacy model names deepseek-chat and deepseek-reasoner retire on July 24, marking the full start of the V4 era.

I. Performance Snapshot: Near Opus-Level, Coding Closing In on GPT-5.6

Early developer feedback says DeepSeek V4 GA punches hard:

• Overall level: approaches Anthropic Opus 4.8; coding ability tracks OpenAI GPT-5.6 Sol closely
• Agent ability: much stronger — 3D rendering and SVG generation improve markedly
• In the wild: one-shot 3D shooters, HTML games mixing Minecraft and No Man's Sky vibes, even classic Cut the Rope running successfully in a single pass

Architecturally, both DeepSeek V4 variants use MoE (Mixture-of-Experts) and ship a default 1M-token context window. Pro totals 1.6T parameters with ~49B activated per forward pass; Flash totals 284B with 13B activated. Whether you are crunching ultra-long financial filings, drafting novels, or building complex Agent workflows, million-token context is a real “cheat-code” capability.

Some developers note that V4 needs more iteration rounds than Fable 5 for the same task, and Pro does not lead Flash by a wide margin. In today's landscape, V4 may not beat the newly released Kimi K3 on raw strength — but its price will be significantly lower.

II. The Price Disruptor Gets a Meter: Peak-Valley Billing Arrives

The biggest change this time is not performance — it is pricing strategy.

DeepSeek introduces peak-valley billing for the first time — peak hours (daily 9:00–12:00 and 14:00–18:00 Beijing time) cost 2× the off-peak rate.

Concrete pricing:

Model	Billable item	Off-peak	Peak
V4 Pro	Output / 1M tokens	$0.87	$1.74
V4 Pro	Input (cache miss)	$0.435	$0.87
V4 Flash	Output / 1M tokens	$0.28	$0.56
V4 Flash	Input (cache hit)	$0.0028	$0.0056

Cross-check: Fable 5 output is $50 / 1M tokens, GPT-5.6 Sol $30, Kimi K3 $15. Even V4 Pro peak pricing is under 3.5% of rivals; Flash is almost rounding error.

Even with a “meter” attached, V4 remains the same price disruptor that clears the board.

Impact is limited for individuals, but teams running Agents 24/7 should replan orchestration — shift batch jobs and evals to off-peak hours to keep spend under control. Cache hits stay extremely cheap; fixed prompts and long-context reuse remain key cost levers.

III. End of an Era: The V4 Era Officially Begins

One more note: deepseek-chat and deepseek-reasoner retire on July 24.

All platform developers must switch to the dedicated IDs deepseek-v4-pro and deepseek-v4-flash or requests will fail — unmigrated pipelines will hard-break. For compatibility, the old names currently map to deepseek-v4-flash non-thinking and thinking modes respectively.

Handoff complete — the V4 era is on.

Urgent for developers: if you still call deepseek-chat or deepseek-reasoner, finish migration before July 24 23:59. Use the DeepSeek API open platform console to switch to V4 model IDs quickly.

IV. Summary

From current signals, V4 is unlikely to be “#1 on every scoreboard” — but DeepSeek has never competed on the performance card alone.

The real story is the path that keeps proving out: Opus-class capability at a fraction of rival pricing.

Whether you try the DeepSeek web app or integrate via the DeepSeek API open platform, V4 GA is worth an immediate look. Top-tier performance, still a price disruptor — can this GA moment recreate another “DeepSeek moment”?

Further Reading

• DeepSeek V4 Official Launch July 15 Explained
• DeepSeek V4 Peak-Valley Pricing Deep Dive
• DeepSeek V4 API Docs