DEEPSEEK V4
GA_FULL_
RELEASE_2026.

DeepSeek V4 GA release architecture benchmarks and peak-valley API pricing

TL;DR: After three months of preview access, DeepSeek V4 officially hit general availability on July 20, 2026. The GA build keeps the same 1.6T MoE architecture but adds production-tuned agent, math, and coding performance — plus something new from DeepSeek: peak-valley API pricing. This guide covers the full timeline, GA vs. preview changes, V4-Pro/Flash architecture (CSA, HCA, mHC, Muon), benchmark tables (SWE-bench Verified 80.6% open-weight record), the 116x cost gap vs Claude Fable 5, comparisons against GPT-5.6 Sol, peak-hour savings tactics, and the July 24 migration deadline for deepseek-chat and deepseek-reasoner.

30-Second Read · TL;DR

GA launch2026-07-20 · preview since April 24 · MIT-licensed V4-Pro + V4-Flash
Top benchmark80.6% SWE-bench Verified · open-weight record · LiveCodeBench 93.5%
Context1M tokens · KV cache 10% of V3.2 (Flash: 7%) · 384K max output
PricingOff-peak output $0.87/M (Pro) · peak doubles · Fable 5 costs 116x more per task
Migration deadline2026-07-24 15:59 UTC · deepseek-chat / deepseek-reasoner retire

1. Timeline: From Preview to Full GA

DateEvent
April 24, 2026V4 preview launches + open-sources under MIT — V4-Pro (1.6T) and V4-Flash (284B)
May 2026Production-tuned V4-Flash and V4-Pro builds ship; API goes live
June 2026V4-Pro output price permanently cut 75% to $0.87/M tokens
June 29, 2026DeepSeek emails all API users: GA mid-July, first disclosure of peak-valley pricing
July 19, 2026Developers receive GA gray-test access; media reports "full release imminent"
July 20, 2026GA officially launches (today)
July 24, 2026Legacy endpoints deepseek-chat and deepseek-reasoner permanently disabled at 15:59 UTC

The preview was already strong. GA is the same architecture graduating into production — with targeted gains in agentic workflows, math reasoning, and code generation, plus a formal commercial pricing structure.

2. What's New in GA vs. the Preview

What ChangedPreview (April 2026)GA (July 2026)
PerformanceBaselineImproved: agentic tasks, math reasoning, code generation
PricingFlat ratePeak/off-peak pricing introduced
Model namesdeepseek-v4-pro, deepseek-v4-flashSame
Legacy endpointsdeepseek-chat routes to V4-FlashRetiring July 24, 2026
AvailabilityPreview (limited)Full GA, globally available

The underlying MoE design has not changed. DeepSeek is keeping the same V4 architecture that shipped in April. The GA upgrade is about stability, optimization, and committing to a disciplined commercial pricing model.

3. Model Family Specs: V4-Pro vs V4-Flash

SpecV4-ProV4-Flash
Total parameters1.6 trillion (1.6T)284 billion (284B)
Active per token49B (3% of total)13B (4.6% of total)
Transformer layers6143
Context window1,000,000 tokens1,000,000 tokens
Max output384K tokens384K tokens
PrecisionFP4 (MoE experts) + FP8 (rest)FP4 + FP8 mixed
Training data33T+ tokens32T+ tokens
LicenseMITMIT

Both models are Mixture-of-Experts (MoE) — only a small fraction of parameters activate per token. That is how you get 1.6T total parameters but only 49B doing actual work on any given inference call.

4. Architecture: CSA, HCA, mHC, and Muon

Most LLMs support 1M token context windows on paper but choke in practice — KV cache memory alone becomes prohibitive. DeepSeek V4 actually solves this. The V4 team replaced Multi-head Latent Attention (MLA) from V2/V3 with a two-track hybrid system.

4.1 Compressed Sparse Attention (CSA)

  • Compresses KV entries 4x via softmax-gated pooling
  • Uses an FP4 "lightning indexer" for top-k sparse selection — top-1024 for Pro, top-512 for Flash
  • Retains a 128-token sliding window for recent context
  • At 1M tokens: only 27% of the inference FLOPs vs. DeepSeek V3.2

4.2 Heavily Compressed Attention (HCA)

  • Compresses tokens 128x, then runs full dense attention over the global cache
  • Captures long-range dependencies CSA might miss
  • V4-Flash uses HCA in the first 2 layers, then alternates CSA/HCA; V4-Pro follows a similar pattern

CSA and HCA alternate across transformer layers (not run in parallel). Combined result at 1M tokens:

  • KV cache drops to 10% of V3.2's memory footprint (V4-Flash: just 7%)
  • That efficiency gain is what makes serving 1M-token contexts economically viable at scale

4.3 Manifold-Constrained Hyper-Connections (mHC)

Standard residual connections (x = x + f(x)) degrade in very deep networks. mHC replaces them with a 4-channel residual stream governed by a doubly stochastic matrix (Birkhoff polytope constraint). Signal propagates stably through 61 layers without sacrificing expressivity — deeper network, more stable training.

4.4 Muon Optimizer

DeepSeek trained V4 with the Muon optimizer instead of AdamW, using Newton-Schulz orthogonalization to condition gradient steps. Result: faster convergence and more stable training runs at scale.

4.5 Three Inference Modes

ModeDescriptionBest For
Non-thinkFast, no chain of thoughtSimple queries, routing, classification
Think HighExplicit reasoning in <redacted_thinking> blocksDebugging, medium-complexity tasks
Think MaxMaximum reasoning effort (requires 384K+ context)Complex math, multi-step agent planning

Recommended sampling for all modes: temperature=1.0, top_p=1.0 (per official docs).

5. Benchmarks: Where V4-Pro Wins and Where It Doesn't

BenchmarkDeepSeek V4-ProClaude Fable 5GPT-5.6 UltraClaude Opus 4.8
SWE-bench Verified80.6% (open-weight record)96.0%N/A~69%
SWE-bench Pro55.4%80.3%78.1%69.2%
LiveCodeBench (Pass@1)93.5%88.1%87.4%83.2%
Codeforces Elo3,206
Terminal-Bench 2.183.9%88.0%85.1%82.7%

Honest takeaways:

  • On SWE-bench Verified (real GitHub bug fixes), V4-Pro at 80.6% is the best open-weight score — tied with Gemini 3.1 Pro, far ahead of any other open model
  • Claude Fable 5 still leads overall: 96% on Verified, 80.3% on the harder SWE-bench Pro
  • V4-Pro beats everyone on LiveCodeBench and Codeforces Elo — the right pick for algorithmic problem-solving
  • The gap narrows dramatically when you factor in cost (see next section)

For routing context on how V4-Flash fits into multi-model pipelines, see our OpenRouter programming leaderboard + V4-Flash multi-route guide and Apple Silicon DeepSeek V4 local benchmark coverage.

6. Cost-Performance: The 116x Gap

Artificial Analysis ran major models through industry-specific indices. On the Strategy & Operations benchmark:

ModelScoreCost per Task
Claude Fable 550 points$3.48
DeepSeek V4-Pro38 points$0.03
DeepSeek V4-FlashAll six index categories: under $0.04 per task

Fable 5 costs 116x more than V4-Pro for a 24% performance difference (50 vs 38 points). For high-volume production workloads, that math is hard to ignore.

7. DeepSeek V4 vs GPT-5.6 Sol vs Claude Fable 5

DimensionDeepSeek V4-ProGPT-5.6 SolClaude Fable 5
Open weights / self-hostableYes (MIT)NoNo
1M contextYesNot confirmedYes
Best overall codingSecond tierSecond tierLeads SWE-bench Pro
Best algorithms/mathLiveCodeBench leaderStrong
Output cost (off-peak)$0.87/M tokens~$15/M~$50/M
Output cost (peak)$1.74/M tokens
Agent capabilityStrong, multi-agent orchestrationStrongStrongest
Data sovereigntyYes (self-host possible)NoNo

When to pick DeepSeek V4-Pro: self-hosting for compliance, high-volume API calls on a budget, coding agents, document analysis, or long-context reasoning.

When to pick Claude Fable 5: absolute best on complex multi-file repository work (80.3% SWE-bench Pro) when budget is not the primary constraint.

When to pick GPT-5.6 Sol: terminal/shell-based agentic workflows (85.1% Terminal-Bench 2.1) or deep Microsoft 365 / Azure Copilot integration (GPT-5.6 became the preferred M365 model on July 9). See our GPT-5.6 Sol Ultra analysis for frontier-model context.

8. Peak-Valley Pricing: Full Table and Savings Tips

This is the most significant change in the GA release. DeepSeek introduced time-based pricing — rates double during Beijing business hours, similar to time-of-use electricity pricing.

8.1 Full Pricing Table

ModelItemOff-Peak RatePeak Rate
V4-ProInput (cache hit)$0.0035 / 1M tokens (¥0.025)2x
Input (cache miss)$0.435 / 1M (¥3.00)$0.87 / 1M (¥6.00)
Output$0.87 / 1M (¥6.00)$1.74 / 1M (¥12.00)
V4-FlashInput (cache hit)$0.0028 / 1M (¥0.02)2x
Input (cache miss)$0.14 / 1M (¥1.00)$0.28 / 1M (¥2.00)
Output$0.28 / 1M (¥2.00)$0.56 / 1M (¥4.00)

Peak hours (Beijing Time): weekdays 09:00–12:00 and 14:00–18:00.

8.2 How to Minimize Your Bill

  1. Schedule batch jobs off-peak. Bulk summarization, data labeling, code review pipelines, and nightly report generation should run after 18:00 Beijing time or before 09:00. A simple cron job cuts your bill in half.
  2. Maximize cache hit rate. Consistent system prompts pay just $0.0028/M on V4-Flash cache hits — essentially free. Put static prompt content first so it gets cached.
  3. Use V4-Flash as a first-pass filter. Route intent classification, entity extraction, and FAQ responses to V4-Flash ($0.28/M output). Escalate complex reasoning to V4-Pro. Tiered routing can cut inference costs 60–80%.
  4. Watch your email. DeepSeek commits to billing change notifications 24 hours in advance. Keep your account email monitored.

Even at peak pricing, 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). The peak surcharge is real, but the baseline remains the best value in the market.

9. API Migration Guide: Deadline July 24, 2026

Warning: Legacy model names deepseek-chat and deepseek-reasoner will be permanently disabled on July 24, 2026 at 15:59 UTC (July 24, 23:59 Beijing time).

9.1 Migration Mapping

Old Model NameNew EquivalentNotes
deepseek-chatdeepseek-v4-flash (non-thinking)Drop-in replacement for most chat use cases
deepseek-reasonerdeepseek-v4-flash (thinking mode)Or upgrade to deepseek-v4-pro for stronger reasoning

9.2 OpenAI SDK (Python)

# Old (stops working July 24) response = client.chat.completions.create( model="deepseek-chat", messages=[{"role": "user", "content": "Hello"}] ) # New (works now and after July 24) response = client.chat.completions.create( model="deepseek-v4-flash", # or "deepseek-v4-pro" messages=[{"role": "user", "content": "Hello"}] ) # Enable thinking mode (replaces deepseek-reasoner behavior) response = client.chat.completions.create( model="deepseek-v4-pro", messages=[{"role": "user", "content": "Solve this step by step..."}], extra_body={ "thinking": {"type": "enabled", "budget_tokens": 8000} } )

9.3 Anthropic SDK

V4 supports both OpenAI-compatible and Anthropic Messages API formats. Swap the model name — the base URL stays the same:

import anthropic client = anthropic.Anthropic( api_key="your-deepseek-api-key", base_url="https://api.deepseek.com" ) message = client.messages.create( model="deepseek-v4-pro", max_tokens=4096, messages=[{"role": "user", "content": "What changed in the V4 GA release?"}] )

9.4 Migration Checklist

  • Search your codebase for deepseek-chat and deepseek-reasoner
  • Update to deepseek-v4-flash or deepseek-v4-pro
  • Decide non-thinking vs. thinking mode per use case
  • Test in staging before July 24
  • For Think Max mode, ensure context window is set to 384K+

10. Frequently Asked Questions

Q1: What changed between preview and GA?
A: Same MoE architecture. GA adds production-tuned performance, peak/off-peak pricing, global availability, and retires legacy model names on July 24.

Q2: What are peak hours?
A: Beijing Time weekdays 09:00–12:00 and 14:00–18:00. Rates double. Off-peak is before 09:00, after 18:00, and weekends.

Q3: How does V4-Pro compare to Claude Fable 5 on coding?
A: V4-Pro leads open-weight models at 80.6% SWE-bench Verified. Fable 5 leads overall at 96% Verified and 80.3% SWE-bench Pro. V4-Pro wins LiveCodeBench at 93.5%.

Q4: What replaces deepseek-chat?
A: Use deepseek-v4-flash in non-thinking mode. For deepseek-reasoner, enable thinking on Flash or upgrade to deepseek-v4-pro.

Q5: Can I self-host V4?
A: Yes — both V4-Pro and V4-Flash are MIT-licensed. V4-Pro needs serious GPU clusters; V4-Flash is more practical for smaller deployments.

Q6: Is V4 still cheaper than GPT-5.6 at peak?
A: Yes. Peak V4-Pro output is $1.74/M vs ~$15/M for GPT-5.6 Sol — roughly 8.6x cheaper.

11. Deep Insight: The Open-Source Milestone That Changes the Economics

DeepSeek V4 GA is not primarily about beating Claude Fable 5 or GPT-5.6 on every leaderboard — it cannot, not yet. Its value is delivering open-weight frontier-class performance at one-tenth to one-hundredth the cost of closed-source alternatives.

The CSA + HCA attention redesign is the engineering story underneath the benchmarks. Replacing MLA with a hybrid that alternates 4x sparse compression and 128x global compression is what turns "1M context on the spec sheet" into "1M context you can actually afford to serve." KV cache at 10% of V3.2 is not an incremental optimization — it is the difference between a demo and a production deployment.

mHC and Muon are quieter but equally important. Training a 61-layer, 1.6T MoE model to convergence without gradient pathology requires architectural discipline, not just more GPUs. The MIT license means enterprises that cannot send data to US APIs — or developers who want to fine-tune on proprietary codebases — finally have a credible open alternative that scores 80.6% on real GitHub bug fixes.

Peak-valley pricing signals maturity, not retreat. DeepSeek is moving from "price disruptor" to "sustainable platform" while keeping off-peak rates aggressively cheap. For teams running 24/7 agent pipelines, the operational answer is straightforward: batch work off-peak, cache system prompts aggressively, and route simple calls to V4-Flash. The 116x cost gap vs Fable 5 means even a 24-point benchmark deficit is often the rational tradeoff at scale.

The July 24 migration deadline is the immediate action item. Four days to grep your repos, update model strings, and validate in staging. The longer-term bet: V4-Pro is the default open-weight foundation for agent engineers, compliance-sensitive enterprises, and anyone building on a budget who needs near-frontier coding performance with a 1M-token window that actually works.

12. Closing: API Works Anywhere — Agent Dev Still Belongs on Mac

Updating your API key, swapping model names in Cursor, and routing through OpenRouter — any laptop handles that. But if you want to run Cursor + V4-Pro long-context coding agents, stress-test Think Max reasoning on multi-repo SWE-bench tasks, or validate quantized V4-Flash weights locally with MLX before production deploy, Apple Silicon unified memory and Metal remain the lowest-friction path.

The practical split: keep your daily driver on API calls, and offload V4 agent regression suites, 1M-token document batch jobs, and peak/off-peak cron scheduling tests to a MACGPU remote Mac mini M4 node — on-demand, SSH-secured. Before the July 24 deadline hits, isolate agent workloads so your primary machine stays stable for everyday development.