What is the deadline to migrate from deepseek-chat and deepseek-reasoner?

Legacy model IDs deepseek-chat and deepseek-reasoner are retired on July 24, 2026 at 15:59 UTC. Map deepseek-chat to deepseek-v4-flash (non-thinking) and deepseek-reasoner to deepseek-v4-flash with thinking enabled or deepseek-v4-pro for maximum reasoning depth.

When do peak-hour API rates apply?

Peak pricing applies on weekdays from 09:00–12:00 and 14:00–18:00 Beijing time (UTC+8). During peak windows every listed rate doubles; off-peak rates match the pre-GA baseline.

Should I use V4-Pro or V4-Flash in production?

Route high-volume Agent loops, batch jobs, and latency-sensitive chat to V4-Flash. Reserve V4-Pro for complex repo-level reasoning, Think Max mode (384K+ output), and tasks where SWE-bench Pro quality matters more than unit cost.

Can I self-host DeepSeek V4 on a Mac?

Yes. Both variants ship under MIT. V4-Flash (~160 GB weights) runs on 128 GB+ Apple Silicon via antirez ds4; V4-Pro (~865 GB) needs a 512 GB Mac Studio or multi-GPU server. See the VPSMAC ds4 guide for memory thresholds and Mac VPS options.

Are the benchmark numbers trustworthy?

Scores such as SWE-bench Verified 80.6% come from DeepSeek's tech report with vendor-specific harnesses. Independent reproductions are ongoing. Treat benchmarks as directional and validate with your own pilot tasks before production routing.

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

Q: How does V4 compare to Kimi K3?

Kimi K3 leads on raw parameter scale (2.8T vs 1.6T) and several long-horizon coding benchmarks, but DeepSeek V4-Pro wins on cost-per-task — roughly 116× cheaper than Claude Fable 5 on comparable SWE workloads — and already has full GA weights plus peak-valley API billing.

If your production stack still calls deepseek-chat or routes everything through a single closed-source API, the July 20 GA of DeepSeek V4 changes the math. This guide for API teams and model buyers covers the full release timeline, V4-Pro vs V4-Flash architecture, benchmark tables against GPT-5.6 Sol and Claude Fable 5, peak-valley pricing, four cost-saving tactics, Python SDK migration code, a five-step Runbook, and six FAQs — everything you need before the July 24 legacy endpoint sunset.

Pain Points: Why V4 GA Forces a Routing Reset

Legacy endpoints die in four days. deepseek-chat and deepseek-reasoner retire on July 24, 2026 at 15:59 UTC. Any cron job, OpenClaw gateway, or CI pipeline still pointing at the old IDs will hard-fail without a config change — and most teams discover this only when the first off-hours batch job breaks.
Peak-valley billing is new and easy to underestimate. GA introduces 2× rates during Beijing weekday peak windows (09:00–12:00 and 14:00–18:00). A 24/7 Agent that cannot shift load sees an effective 20–30% bill increase even though off-peak baseline prices stayed flat after June's 75% cut.
Model selection got harder, not simpler. V4 ships two MIT-licensed variants plus three reasoning modes, while Kimi K3 now holds the open-weight scale record at 2.8T parameters. Teams need a routing matrix — not a single default model — to capture V4's 116× cost advantage over Fable 5 without sacrificing quality on hard repo tasks.

Release Timeline: Preview to GA in 87 Days

Date	Milestone
April 24, 2026	V4 Preview launch — V4-Pro and V4-Flash under MIT, 1M context, CSA/HCA architecture unveiled
May 2026	Production tuning — concurrency scaled to 500+ for V4-Pro, ds4 Metal engine ships for Mac local inference
June 2026	Permanent 75% price cut on V4-Pro API rates; OpenRouter weekly volume for V4-Flash crosses 3.43T tokens
June 29, 2026	Email to API users announcing mid-July GA and upcoming peak-valley pricing mechanism
July 19, 2026	Gray release — select accounts receive GA endpoints early for validation
July 20, 2026	General Availability (GA) — official V4 with peak-valley billing and performance optimizations
July 24, 2026	Legacy retirement — `deepseek-chat` and `deepseek-reasoner` deprecated at 15:59 UTC

What Is DeepSeek V4?

DeepSeek V4 is a dual-variant MoE family — V4-Pro for maximum reasoning depth and V4-Flash for high-throughput Agent workloads — both shipping with a 1M-token context window, 384K max output, and an MIT license. Unlike V3, which split thinking and non-thinking into separate model IDs, V4 exposes reasoning effort as a request parameter across a single weight set.

Spec	V4-Pro	V4-Flash
Total parameters	1.6T	284B
Active per token	49B	13B
Transformer layers	61	43
Context window	1,048,576 tokens (1M)
Max output	384,000 tokens
Precision	FP4 weights + FP8 activations
Training data	33T+ tokens	32T+ tokens
License	MIT
API model ID	`deepseek-v4-pro`	`deepseek-v4-flash`
Weights on disk	~865 GB	~160 GB

One-liner: DeepSeek V4 is an open-weight, 1M-context coding AI that matches frontier closed models on SWE-bench Verified at a fraction of the price — with GA peak-valley billing and a hard July 24 migration deadline.

Architecture: CSA, HCA, mHC & Muon

V4's headline is not raw parameter count — it's serving cost reduction. The attention stack combines two compression mechanisms plus training innovations that cut FLOPs and KV memory versus V3.2.

Compressed Sparse Attention (CSA)

4× compression on attention maps via learned sparsity patterns;
FP4 lightning indexer selects relevant KV blocks at near-zero overhead;
Retains top-1024 keys for Pro and top-512 for Flash per head;
128-token sliding window preserves local coherence between sparse blocks;
Net result: 27% of V3.2 FLOPs at equivalent context length.

Heavily Compressed Attention (HCA)

128× compression on selected layers via low-rank KV projection;
Deployed on alternating layers — CSA handles precision, HCA handles bulk memory;
KV cache footprint drops to 10% of V3.2 for Pro and 7% for Flash.

Manifold-Constrained Hyper-Connections (mHC)

Standard residual connections accumulate noise across depth. mHC routes information through a 4-channel residual pathway constrained to the Birkhoff polytope — ensuring each channel's mixing weights form a doubly-stochastic matrix. This stabilizes training at 60+ layers without the representation collapse seen in naive deep MoE stacks.

Muon Optimizer

V4 continues DeepSeek's use of the Muon optimizer (Momentum Updated Orthogonalized) for large-matrix parameters, paired with AdamW on embeddings and norms. Moonshot's Kimi K3 independently adopted per-head Muon variants — convergent evidence that orthogonalized updates scale better for trillion-parameter MoE training.

Three Reasoning Modes

V4 exposes reasoning depth as a request parameter rather than a separate model:

Mode	API parameter	Behavior	Best for
Non-think	`thinking: disabled`	Direct response, lowest latency and cost	Classification, summarization, simple codegen
Think High	`thinking: enabled`	Extended chain-of-thought before answer	Multi-step debugging, math, Agent planning
Think Max	`thinking: max`	Deepest reasoning, supports 384K+ output	Long-form proofs, repo-wide refactors, research

Official recommendation for thinking modes: set temperature=1.0 and top_p=1.0. Lower values can collapse diversity in extended reasoning chains.

Benchmark Results

Benchmark	V4-Pro (GA)	V4-Flash (GA)
SWE-bench Verified	80.6%	79.0%
SWE-bench Pro	55.4%	—
LiveCodeBench	93.5%	91.6%
Codeforces Elo	3,206	3,052
Terminal-Bench 2.1	83.9%	—

Reading the table: V4-Pro's 80.6% on SWE-bench Verified places it within striking distance of Claude Opus 4.8 territory. Terminal-Bench 2.1 at 83.9% signals strong tool-use and shell interaction — relevant for OpenClaw and CI Agent deployments. Codeforces Elo 3,206 confirms competition-grade algorithmic skill.

⚠️ Caveat: These are DeepSeek self-reported scores with vendor-specific harnesses. Independent third-party reproductions are still in progress.

Cost-Performance: 116× Cheaper Than Fable 5

Benchmark points alone miss the production story. DeepSeek published a task-level cost comparison on SWE-style workloads:

Model	Cost per task	Quality score	Notes
Claude Fable 5	$3.48	50 pts	Frontier quality, highest unit cost
DeepSeek V4-Pro	$0.03	38 pts	116× cheaper; 76% of Fable quality at <1% cost

For high-volume Agent loops processing thousands of tasks daily, this gap dominates total cost of ownership more than any single benchmark delta.

Head-to-Head: V4-Pro vs GPT-5.6 Sol vs Claude Fable 5

Dimension	DeepSeek V4-Pro	GPT-5.6 Sol	Claude Fable 5
SWE-bench Verified	80.6%	~78%	~82%
Terminal-Bench 2.1	83.9%	88.8%	84.6%
LiveCodeBench	93.5%	~91%	~90%
Context window	1M	400K	200K
Output $/1M (off-peak)	$0.87	~$30	~$50
Open weights (MIT)	Yes	No	No
Self-host option	Yes (865 GB)	No	No
Export-control risk	Low (Chinese vendor, MIT weights)	Medium	High (Fable 5 access restrictions)

V4-Pro does not win every row — GPT-5.6 Sol still leads Terminal-Bench and Fable 5 edges SWE-bench Verified — but the price/context/open-weight combination is unmatched among frontier-tier models.

Which Model When: Scenario Matrix

Scenario	Recommended model	Why
High-volume Agent / batch loops	V4-Flash (Non-think)	Lowest $/token; 13B active params still strong on LiveCodeBench
Multi-step coding Agent	V4-Flash (Think High)	Good reasoning at ~3× lower cost than Pro
Complex repo-level bug fixes	V4-Pro (Think Max) or Claude Fable 5	SWE-bench Pro 55.4%; Fable 5 still leads FrontierSWE
Terminal / shell-heavy workflows	GPT-5.6 Sol	Terminal-Bench 2.1 leader at 88.8%
1M-token document analysis	V4-Pro or Kimi K3	Both offer 1M context; K3 leads some long-doc benchmarks
Cost-sensitive production default	V4-Flash	$0.28/M output off-peak; cache hit input at $0.0028/M
Self-hosted / air-gapped deploy	V4-Flash via ds4 on Mac	160 GB weights fit 128 GB Mac; MIT license
Maximum reasoning depth	V4-Pro (Think Max)	384K+ output; 49B active params per token

Peak-Valley Pricing: Full Rate Tables

GA introduces peak-valley billing — the first major LLM API to price by time-of-day. Off-peak rates match the June post-discount baseline. Peak rates are exactly 2× off-peak.

Peak hours (Beijing time, UTC+8): weekdays 09:00–12:00 and 14:00–18:00. All other times — including nights, weekends, and holidays — bill at off-peak rates.

V4-Pro — USD (per 1M tokens)

Billing item	Off-peak	Peak (2×)
Input (cache hit)	$0.003625	$0.00725
Input (cache miss)	$0.435	$0.87
Output	$0.87	$1.74

V4-Pro — CNY (per 1M tokens)

Billing item	Off-peak (¥/M)	Peak (¥/M)
Input (cache hit)	¥0.025	¥0.05
Input (cache miss)	¥3.00	¥6.00
Output	¥6.00	¥12.00

V4-Flash — USD (per 1M tokens)

Billing item	Off-peak	Peak (2×)
Input (cache hit)	$0.0028	$0.0056
Input (cache miss)	$0.14	$0.28
Output	$0.28	$0.56

V4-Flash — CNY (per 1M tokens)

Billing item	Off-peak (¥/M)	Peak (¥/M)
Input (cache hit)	¥0.02	¥0.04
Input (cache miss)	¥1.00	¥2.00
Output	¥2.00	¥4.00

Four Cost-Saving Tips

Schedule batch jobs off-peak. Queue SWE-style refactors, test generation, and document summarization for Beijing nights and weekends. A job that costs $87 per 100M output tokens off-peak costs $174 during peak.
Structure prompts for cache hits. Keep a static system prompt prefix across requests. Cache-hit input on V4-Flash is $0.0028/M — roughly 1% of cache-miss cost. Agent frameworks with fixed tool schemas benefit most.
Route by task complexity. Default to V4-Flash Non-think for classification and simple codegen. Escalate to V4-Pro Think Max only when SWE-bench Pro-level reasoning is required. Reserve Fable 5 for tasks where V4 trails by >10 points.
Disable thinking when you don't need it. Think High and Think Max consume hidden reasoning tokens billed as output. For latency-sensitive chat and RAG Q&A, Non-think mode cuts both cost and time-to-first-token dramatically.

API Migration: Model ID Mapping & Code Examples

Legacy endpoints map cleanly to V4:

Legacy ID	Replacement	Notes
`deepseek-chat`	`deepseek-v4-flash` (Non-think)	Drop-in for chat/completion workloads
`deepseek-reasoner`	`deepseek-v4-flash` (Think High) or `deepseek-v4-pro`	Use Pro for maximum reasoning depth

Hard deadline: July 24, 2026 at 15:59 UTC. After this timestamp, legacy IDs return errors.

OpenAI SDK (Python)

from openai import OpenAI

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

# V4-Flash with Think High reasoning
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[
        {"role": "system", "content": "You are a senior software engineer."},
        {"role": "user", "content": "Refactor this module for async I/O..."}
    ],
    temperature=1.0,
    top_p=1.0,
    extra_body={"thinking": {"type": "enabled"}}
)

# V4-Pro Think Max for deep reasoning
response_pro = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Analyze this entire repo for security issues."}],
    temperature=1.0,
    top_p=1.0,
    extra_body={"thinking": {"type": "max"}}
)

Anthropic SDK (Python)

import anthropic

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

message = client.messages.create(
    model="deepseek-v4-pro",
    max_tokens=8192,
    temperature=1.0,
    messages=[
        {"role": "user", "content": "Review this pull request for logic errors and suggest fixes."}
    ]
)

print(message.content[0].text)

Both SDK paths hit the same V4 weights. Pick OpenAI format for existing LangChain/LiteLLM integrations; Anthropic format if your stack already uses Claude-compatible clients.

Citable Technical Facts (EEAT)

Attention efficiency: CSA + HCA reduce inference FLOPs to 27% of V3.2; KV cache memory is 10% of V3.2 (Pro) and 7% (Flash).
Cost-per-task: V4-Pro at $0.03/task vs Claude Fable 5 at $3.48/task — a 116× unit-cost gap on SWE-style workloads.
Context/output ceiling: 1M input and 384K max output on both variants — Think Max can consume the full output budget for extended proofs.
Training scale: 33T+ tokens (Pro) and 32T+ tokens (Flash) with Muon optimizer and mHC 4-channel residuals.
Peak-valley spread: Off-peak V4-Flash output at $0.28/M vs peak at $0.56/M — schedule-aware routing saves up to 50% on time-sensitive batches.

Five-Step Migration Runbook

Step 1  Audit — grep all repos/configs for deepseek-chat and deepseek-reasoner; list every service, cron, and Agent gateway route

Step 2  Map — deepseek-chat → deepseek-v4-flash (Non-think); deepseek-reasoner → deepseek-v4-flash Think High or deepseek-v4-pro Think Max

Step 3  Configure — set base_url to https://api.deepseek.com (OpenAI) or /anthropic (Anthropic SDK); temperature=1.0, top_p=1.0 for thinking modes

Step 4  Pilot — run 20 representative tasks off-peak AND during a peak window; log tokens, cache hits, quality, and $/task

Step 5  Deploy hybrid routing — bulk Agent → V4-Flash off-peak; hard reasoning → V4-Pro Think Max; fallback Fable 5 / GPT-5.6 Sol where V4 trails; cut over before Jul 24 15:59 UTC

Frequently Asked Questions

What is the deadline to migrate from deepseek-chat?

July 24, 2026 at 15:59 UTC. After that, deepseek-chat and deepseek-reasoner return API errors. Map them to deepseek-v4-flash and deepseek-v4-pro respectively.

When do peak-hour rates apply?

Weekdays 09:00–12:00 and 14:00–18:00 Beijing time (UTC+8). Every listed rate doubles during peak. Nights, weekends, and holidays bill at off-peak baseline.

Should I use V4-Pro or V4-Flash?

V4-Flash for volume Agent loops and cost-sensitive defaults. V4-Pro for SWE-bench Pro-level reasoning, Think Max mode, and tasks where the extra 36B active parameters matter.

Can I self-host V4 on a Mac?

Yes — both variants are MIT-licensed. V4-Flash runs on 128 GB+ Apple Silicon via antirez ds4. V4-Pro needs 512 GB unified memory or a multi-GPU server.

How does V4 compare to Kimi K3?

Kimi K3 leads on parameter scale (2.8T) and several long-horizon benchmarks, but V4-Pro wins decisively on cost-per-task and already has GA API endpoints with peak-valley billing.

Are the benchmarks trustworthy?

Self-reported with vendor harnesses. SWE-bench Verified 80.6% is directionally credible but should be validated with your own pilot before production routing decisions.

Bottom Line

DeepSeek V4 GA is not a cosmetic version bump — it is a production-grade open-weight family with 1M context, MIT licensing, frontier-near coding benchmarks, and a pricing model that undercuts Claude Fable 5 by orders of magnitude. The peak-valley mechanism adds scheduling complexity, but off-peak rates stayed flat after June's 75% cut, and teams that batch intelligently pay no premium at all.

Key dates: July 20 → GA live · July 24, 15:59 UTC → legacy endpoints retired.

API-only deployments still leave you without a co-located environment for 7×24 Agent loops, Apple toolchain tasks (Xcode, Fastlane, notarytool), or air-gapped inference. Running ds4 locally on a laptop means sleep-interrupted long jobs and commingled API keys with production repos. For teams that want V4 Flash or Pro self-hosted alongside iOS CI or OpenClaw gateways, renting a VPSMAC M4 Mac cloud node — native macOS, SSH + launchd for unattended agents, elastic 128/256/512 GB memory tiers — is usually more stable than personal hardware or a generic Linux VPS. See our ds4 + DeepSeek V4 on Mac guide for memory thresholds and a reproducible Runbook.

Sources: DeepSeek API pricing docs · antirez ds4 repository · DeepSeek V4 tech report · OpenRouter weekly token rankings

Data as of 2026-07-20. Benchmarks are DeepSeek self-reported. Pricing and model capabilities may update — verify latest official docs before production decisions.