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

About 14 min read · MACCOME · Last updated: July 20, 2026

Who this is for: AI developers, indie hackers, and engineering leads running production API workloads. On July 20, 2026, DeepSeek V4 graduated from preview to general availability (GA) — same 1.6T MoE architecture, but with agent/math/code upgrades and the first peak-valley pricing scheme in DeepSeek history. What you get: full timeline from April preview to GA, CSA/HCA/mHC/Muon architecture breakdown, benchmark tables vs. Claude Fable 5 and GPT-5.6 Sol, the 116x cost-performance gap, peak/off-peak rate tables, four savings tactics, and a July 24 migration runbook for retiring deepseek-chat. Context: local inference options in the ds4 / V4 Flash Mac decision guide; broader routing in the OpenRouter multi-model routing matrix.

bolt

TL;DR — 30-second verdict

  • GA is production-ready V4, not a new model: same V4-Pro (1.6T) and V4-Flash (284B) MoE family, with tuned agent, math, and coding performance.
  • Best open-weight SWE-bench Verified score: 80.6% on V4-Pro — tied with Gemini 3.1 Pro, ahead of every other open model.
  • 116x cheaper than Claude Fable 5 on Artificial Analysis Strategy & Ops tasks ($0.03 vs $3.48 per task) with a 24% score gap.
  • Peak-valley pricing is new: rates double during Beijing weekday peaks (09:00–12:00, 14:00–18:00); off-peak V4-Pro output stays at $0.87/M.
  • Migration deadline July 24, 15:59 UTC: deepseek-chat and deepseek-reasoner go offline — update to deepseek-v4-flash or deepseek-v4-pro now.

Six Pain Points: What Does the GA Release Actually Change?

After three months of preview access, most teams hit the same six decision blind spots when GA landed — not because V4 is unknown, but because the commercial packaging shifted:

  1. Preview vs. GA confusion: GA is not a new architecture. The MoE design from April is unchanged; GA adds production tuning and peak-valley billing, not a parameter bump.
  2. Peak pricing math: Off-peak rates match the June permanent 75% discount ($0.87/M output on V4-Pro). Peak doubles to $1.74/M — still 8.6x cheaper than Claude Opus 4.8, but teams running 24/7 agents need to schedule around Beijing business hours.
  3. Benchmark trust gap: V4-Pro leads open-weight SWE-bench Verified at 80.6%, but Claude Fable 5 hits 96.0% Verified and 80.3% Pro. "Best open source" is not "best overall."
  4. Legacy endpoint risk: deepseek-chat and deepseek-reasoner still work today but die on July 24 at 15:59 UTC. Any cron job or production service still using old model IDs will hard-fail.
  5. Three reasoning modes, one default: Non-think, Think High, and Think Max trade speed for depth. Think Max needs 384K+ context — misconfiguring this burns tokens without quality gains.
  6. Self-host vs. API: MIT-licensed weights are available, but 1.6T MoE inference needs serious hardware. The ds4 local inference guide covers when a 128GB Mac cloud node beats API bills.

The sections below resolve each blind spot with official docs, benchmark tables, and migration code.

What Changed From Preview to GA?

DeepSeek V4 preview shipped April 24, 2026 under MIT license. Today's GA release is that same model family graduating into production — with performance refinements and a formal pricing structure.

DateEvent
April 24, 2026V4 preview + open weights (MIT): V4-Pro (1.6T) and V4-Flash (284B)
May 2026Production-tuned V4-Flash and V4-Pro; API generally available
June 2026V4-Pro permanent 75% price cut — output locked at $0.87/M tokens (off-peak baseline)
June 29, 2026Email to all API users: GA mid-July, first disclosure of peak-valley pricing
July 19, 2026GA gray-release to select developers; media reports "full version imminent"
July 20, 2026GA official launch (today)
July 24, 2026Legacy names deepseek-chat and deepseek-reasoner permanently retired (15:59 UTC)

One-line summary: V4 preview was already strong. GA adds agent capability, math reasoning, and code generation refinements on top, plus the first disciplined commercial billing model DeepSeek has ever shipped.

How Does DeepSeek V4 Make 1M Context Economically Viable?

Most models advertise 1M tokens on paper but choke in production — KV cache memory grows linearly with context length. DeepSeek V4 replaces the MLA design from V2/V3 with a hybrid attention stack that actually cuts serving costs.

Model family specs

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
Pretraining data33T+ tokens32T+ tokens
LicenseMITMIT

Compressed Sparse Attention (CSA)

CSA compresses KV sequences 4x via softmax-gated pooling, then uses an FP4 "lightning indexer" for top-k sparse selection (top-1024 on Pro, top-512 on Flash). A 128-token sliding window preserves recent context. At 1M tokens, V4-Pro uses only 27% of the inference FLOPs compared to DeepSeek V3.2.

Heavily Compressed Attention (HCA)

HCA compresses tokens 128x and runs dense global attention over the compressed cache — capturing long-range dependencies CSA might miss. V4-Flash uses HCA in its first two layers, then alternates CSA/HCA; V4-Pro follows a similar pattern.

Combined effect: KV cache memory drops to 10% of V3.2 on V4-Pro (7% on V4-Flash). That is what makes 1M-token serving economically viable at API scale.

Manifold-Constrained Hyper-Connections (mHC)

Standard residual connections (x = x + f(x)) degrade signal flow in very deep networks. mHC replaces them with a 4-channel residual stream governed by a doubly stochastic matrix (Birkhoff polytope constraint). Result: stable gradient propagation through 61 layers without sacrificing expressivity.

Muon optimizer

V4 was trained with the Muon optimizer instead of AdamW. Newton-Schulz orthogonalization conditions gradient steps, yielding faster convergence and more stable training runs at MoE scale.

Which Reasoning Mode Should You Use?

DeepSeek V4 ships three inference modes. Official recommended sampling for all: temperature=1.0, top_p=1.0.

ModeBehaviorBest for
Non-thinkNo chain of thought; fastest responseSimple Q&A, routing, classification
Think HighExplicit reasoning in <redacted_thinking> blocksMedium complexity tasks, code debugging
Think MaxMaximum reasoning effort; needs 384K+ contextComplex math, long-chain agent planning

Where Does V4-Pro Win — and Where Does It Fall Short?

Core benchmark data from DeepSeek official docs and independent aggregators (each vendor used its own inference harness where noted):

BenchmarkDeepSeek V4-ProClaude Fable 5GPT-5.6 UltraClaude Opus 4.8
SWE-bench Verified80.6%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%

Key takeaways:

  • SWE-bench Verified (real GitHub bug fixes): V4-Pro at 80.6% is the best open-weight score — tied with Gemini 3.1 Pro.
  • SWE-bench Pro (harder variant): Claude Fable 5 leads at 80.3%; V4-Pro at 55.4% is second tier on the hardest repo work.
  • LiveCodeBench + Codeforces: V4-Pro leads competitive programming — relevant if algorithmic problem-solving is your core workload.
  • Terminal-Bench 2.1: Fable 5 (88.0%) and GPT-5.6 (85.1%) beat V4-Pro (83.9%) on shell-heavy agent tasks.
warning

Caveat: Benchmarks are vendor-reported unless independently reproduced. Cross-model harness differences (DeepSeek vs. Codex vs. Claude Code) mean treat scores as directional, not definitive. For live OpenRouter volume context, see the June 2026 OpenRouter rankings analysis.

Is the 116x Cost Gap Real?

Raw benchmark scores tell half the story. Artificial Analysis ran major models through industry-specific task indices. On the Strategy & Operations benchmark:

  • Claude Fable 5: 50 points, $3.48 per task
  • DeepSeek V4-Pro: 38 points, $0.03 per task
  • DeepSeek V4-Flash: all six index categories, each under $0.04 per task

That is 116x cheaper for a roughly 24% score gap (31% relative). For high-volume production pipelines — nightly code review, bulk document summarization, agent loops — the math is hard to ignore even if Fable 5 wins on the hardest single-repo fixes.

DeepSeek V4 vs. GPT-5.6 Sol vs. Claude Fable 5: Which Should You Pick?

DimensionDeepSeek V4-ProGPT-5.6 SolClaude Fable 5
Open weights / MITYesNoNo
Self-hostableYesNoNo
Context window1M tokensNot publicly confirmed1M tokens
Best overall codingSecond tier (80.6% Verified)Second tier (Terminal-Bench 85.1%)Leads (96.0% Verified, 80.3% Pro)
Best algorithms / mathLeads (LiveCodeBench 93.5%)Strong
Output cost (off-peak)$0.87/M tokens~$15/M~$50/M
Output cost (peak)$1.74/M tokens
Data sovereigntyYes (self-host)NoNo
Agent orchestrationStrong; multi-agent supportStrongStrongest

Routing guidance:

  • Budget-limited / high-frequency / private deployment: V4-Pro or V4-Flash
  • Hardest multi-file repo fixes, cost secondary: Claude Fable 5
  • Terminal/shell agent workflows, M365/Azure stack: GPT-5.6 Sol
  • Massive log/document processing at lowest cost: V4-Flash with cache-hit input at $0.0028/M

For how Kimi K3 now stacks up against V4-Pro on parameter count and context, see the Kimi K3 open-source review.

What Does Peak-Valley Pricing Cost — and When?

GA's most debated change: time-based pricing where rates double during Beijing weekday peak hours. Think of it like time-of-use electricity pricing.

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

ModelItemOff-PeakPeak
V4-ProInput (cache hit)$0.0035 / 1M tokens2x
Input (cache miss)$0.435 / 1M tokens$0.87 / 1M
Output$0.87 / 1M tokens$1.74 / 1M
V4-FlashInput (cache hit)$0.0028 / 1M tokens2x
Input (cache miss)$0.14 / 1M tokens$0.28 / 1M
Output$0.28 / 1M tokens$0.56 / 1M

Even at peak, V4-Pro output ($1.74/M) is 8.6x cheaper than Claude Opus 4.8 ($15/M) and GPT-5.6 Sol (~$15/M). The surcharge is real; the baseline remains the market's best value.

Four ways to cut your bill

  1. Schedule batch jobs off-peak. Document processing, data labeling, code review pipelines, and nightly report generation should run after 18:00 Beijing time or before 09:00. A simple cron cut can halve inference spend.
  2. Maximize cache hit rate. Static system prompts cached on V4-Flash cost $0.0028/M input — essentially free. Put stable prompt prefixes first so they get cached across requests.
  3. Route simple tasks to V4-Flash. Intent classification, entity extraction, and FAQ routing on Flash ($0.28/M output); escalate only complex reasoning to V4-Pro. Tiered routing typically cuts costs 60–80%.
  4. Watch your account email. DeepSeek commits to billing change notifications 24 hours in advance. Ensure the registered address is monitored.

Historical pricing context: the June permanent discount is covered in our OpenRouter June 2026 analysis, where V4-Flash already dominated daily token volume.

How Do You Migrate Before July 24?

Deadline: July 24, 2026, 15:59 UTC (July 24, 23:59 Beijing time). Legacy model names deepseek-chat and deepseek-reasoner will permanently stop responding.

Old model nameNew equivalentNotes
deepseek-chatdeepseek-v4-flash (non-thinking)Drop-in for most chat use cases
deepseek-reasonerdeepseek-v4-flash (thinking mode)Or upgrade to deepseek-v4-pro for stronger reasoning

OpenAI SDK (Python):

python
from openai import OpenAI

client = OpenAI(
    api_key="your_deepseek_api_key",
    base_url="https://api.deepseek.com"
)

# Old (stops working July 24)
# response = client.chat.completions.create(
#     model="deepseek-chat",
#     messages=[{"role": "user", "content": "Hello"}]
# )

# New — non-thinking (replaces deepseek-chat)
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Hello"}]
)

# New — thinking mode (replaces deepseek-reasoner)
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}
    }
)

Anthropic SDK (Python): V4 supports both OpenAI ChatCompletions and Anthropic Messages formats. Base URL unchanged; update the model name only.

python
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?"}]
)

Six-Step Runbook: From Evaluation to Production Routing

  1. Audit legacy model IDs. Grep your codebase and CI configs for deepseek-chat and deepseek-reasoner. Any hit must migrate before July 24, 15:59 UTC.
  2. Map workloads to Pro vs. Flash. Simple routing and classification on Flash; complex reasoning and long-chain agents on Pro. Use the comparison table above as your routing matrix.
  3. Pick a reasoning mode per endpoint. Non-think for latency-sensitive paths; Think High for debugging; Think Max only when 384K+ context is configured.
  4. Configure peak-valley scheduling. Move batch jobs to off-peak Beijing hours. Set cron or queue delays so document pipelines and code review runs avoid 09:00–12:00 and 14:00–18:00 weekdays.
  5. Enable prompt caching. Structure system prompts with static prefixes first. Target 90%+ cache hit rates on repeated agent loops — V4-Flash cache-hit input at $0.0028/M is near-zero marginal cost.
  6. Deploy hybrid routing on a persistent node. Route bulk work to V4-Flash, critical repo fixes to Claude Fable 5, terminal agents to GPT-5.6 Sol. Run the Gateway on a dedicated host, not a laptop — see the MACCOME bridge below.

Three Hard Numbers Worth Citing

  • 80.6% — V4-Pro SWE-bench Verified; best open-weight score, tied with Gemini 3.1 Pro
  • 116x — cost gap vs. Claude Fable 5 on Artificial Analysis Strategy & Ops tasks ($0.03 vs $3.48 per task)
  • 10% / 27% — KV cache memory vs. V3.2 (10% on Pro, 7% on Flash) and inference FLOPs (27% of V3.2 at 1M context)

Wrap-Up: Production-Grade Open Weights, Not a Universal Replacement

DeepSeek V4 GA is one of the most important open-model milestones of 2026. CSA/HCA attention, mHC, and Muon training are genuine engineering — not parameter inflation. At roughly one-tenth to one-hundredth the cost of closed frontier models, V4-Pro delivers the strongest open-weight coding performance on SWE-bench Verified, real 1M context, and MIT-licensed weights you can self-host.

It does not beat Claude Fable 5 on the hardest repo work, and GPT-5.6 Sol still leads terminal-heavy agents. Peak-valley pricing adds operational complexity for 24/7 pipelines. But for teams where cost, data sovereignty, or long-context agents matter, V4-Pro remains the default evaluation target.

Running DeepSeek V4 agent workflows in production — OpenClaw Gateway, multi-model hybrid routing, or peak-aware batch schedulers — on a local MacBook hits three structural bottlenecks:

  • Sleep and network jitter: lid-close or Wi-Fi handoff interrupts long agent sessions; consumed tokens are not refunded;
  • Compute contention: local IDE, Simulator, and agents compete for unified memory, slowing long-context throughput;
  • No true 24/7 routing node: hybrid model strategies need a persistent Gateway dispatching by task type — a laptop is a poor scheduler.

For stable V4-Pro + multi-model agent stacks, MACCOME Mac cloud hosts provide real macOS, SSH handoff, and isolated environments so agents run 24/7 on dedicated nodes. Public plans: Mac mini cloud rental rates.

Dates to bookmark: July 20 (GA live) → July 24, 15:59 UTC (legacy endpoints retired).

Sources: DeepSeek official docs, arXiv:2606.19348, HuggingFace Model Hub, LLMReference, Artificial Analysis, MangoMind Blog. Benchmarks as reported July 20, 2026.

FAQ

What is DeepSeek peak-valley pricing?

GA introduces time-based rates that double during Beijing weekday peaks (09:00–12:00 and 14:00–18:00). Off-peak V4-Pro output is $0.87/M; peak is $1.74/M. V4-Flash off-peak output is $0.28/M; peak is $0.56/M.

How do I migrate from deepseek-chat before July 24?

Replace deepseek-chat with deepseek-v4-flash (non-thinking) and deepseek-reasoner with deepseek-v4-flash thinking mode or deepseek-v4-pro. Legacy names stop working July 24, 2026 at 15:59 UTC (23:59 Beijing time).

Is DeepSeek V4 better than GPT-5.6 Sol?

Depends on the task. GPT-5.6 Sol leads Terminal-Bench 2.1 (85.1% vs 83.9%). V4-Pro leads LiveCodeBench (93.5%) and matches the best open-weight SWE-bench Verified score (80.6%). On cost, V4-Pro output at $0.87/M is roughly 17x cheaper than GPT-5.6 Sol at ~$15/M.

What are the three DeepSeek V4 reasoning modes?

Non-think (fast, no chain of thought), Think High (explicit reasoning for medium tasks), and Think Max (maximum effort, needs 384K+ context). Recommended sampling: temperature=1.0, top_p=1.0 for all modes.

Can I self-host DeepSeek V4?

Yes — both V4-Pro (1.6T) and V4-Flash (284B) ship under MIT with 1M context. Production self-hosting needs high-memory GPU clusters. For Mac-based local inference, see the ds4 / V4 Flash decision guide.

How do I run DeepSeek V4 agents 24/7 in production?

Deploy agent gateways on a dedicated Mac cloud host to avoid laptop sleep interrupting long sessions. See MACCOME Mac cloud rental plans for node configs and pricing.