deepseek-chat, need to weigh cost against GPT-5.6 and Claude Fable 5, or want to finish API migration before the July 24 cutoff, this guide covers: the full preview-to-GA timeline; how CSA+HCA architecture sustains a 1M-token context; benchmark scores, peak-valley price tables, and a six-step migration runbook.
What Did DeepSeek V4 GA Ship Today? Preview to Full Release Timeline
On July 20, 2026, DeepSeek V4 reached General Availability. The underlying MoE design is not new — it was open-sourced under MIT on April 24 — but GA is the graduation release: stability hardening, performance tuning, and a formal commercial billing stack. The full release doubles down on Agent workflows, mathematical reasoning, and code generation while introducing disciplined pricing, marking DeepSeek's shift from near-free experimentation to sustainable production economics.
| Date | Milestone |
|---|---|
| 2026-04-24 | V4 preview + MIT open source: V4-Pro (1.6T) and V4-Flash (284B) |
| 2026-05 | Production-tuned V4-Flash and V4-Pro; API generally available |
| 2026-06 | V4-Pro output price permanently cut 75% to $0.87/M tokens |
| 2026-06-29 | Email to all API users announcing mid-July GA and first disclosure of peak-valley pricing |
| 2026-07-19 | Selected developers receive GA gray-release access |
| 2026-07-20 | GA production release (publication date of this article) |
| 2026-07-24 | Legacy endpoints deepseek-chat and deepseek-reasoner permanently retired |
Migration urgency: Legacy model names deepseek-chat / deepseek-reasoner stop responding at 23:59 Beijing time on July 24. Production systems have four days to switch.
Billing complexity: GA introduces peak-valley pricing for the first time. Weekday rates double during 09:00–12:00 and 14:00–18:00 Beijing time — 24/7 pipelines need rescheduling.
Performance expectations: GA does not beat Claude Fable 5 or GPT-5.6 Ultra across the board, but delivers the strongest open-source performance at 1/10 to 1/100 the cost.
Self-hosting barrier: Despite MIT licensing, V4-Pro's 1.6T parameter footprint demands serious hardware. Most teams stay on API or high-spec cloud Mac inference nodes.
Kimi K3 competition: Kimi K3 chases scale records at 2.8T parameters; DeepSeek defends its developer base with pricing and a mature API ecosystem.
DeepSeek V4 Architecture Breakdown and Full Benchmark Table
DeepSeek V4 moves past V2/V3's Multi-head Latent Attention (MLA) to a hybrid of CSA (Compressed Sparse Attention) and HCA (Heavily Compressed Attention). At 1M tokens, KV cache memory drops to just 10% of V3.2 (7% on V4-Flash), and inference FLOPs fall to 27% of V3.2. mHC (manifold-constrained hyper-connections) stabilizes the 61-layer deep network across four residual channels; training uses the Muon optimizer instead of AdamW.
| Spec | V4-Pro | V4-Flash |
|---|---|---|
| Total parameters | 1.6 trillion (1.6T) | 284 billion (284B) |
| Active per token | 49 billion (49B) | 13 billion (13B) |
| Transformer layers | 61 | 43 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output | 384K tokens | 384K tokens |
| Precision | FP4 (experts) + FP8 | FP4 + FP8 mixed |
| Pretraining data | 33T+ tokens | 32T+ tokens |
| License | MIT | MIT |
Three reasoning modes ship out of the box: Non-think (fastest, no chain-of-thought), Think High (explicit reasoning, ideal for code debugging), and Think Max (maximum reasoning depth, needs 384K+ context for complex math and long-chain Agents). Official sampling defaults: temperature=1.0, top_p=1.0.
| Benchmark | DeepSeek V4-Pro | Claude Fable 5 | GPT-5.6 Ultra | Claude Opus 4.8 |
|---|---|---|---|---|
| SWE-bench Verified | 80.6% ★ open-source best | 96.0% | not separately published | ~69% |
| SWE-bench Pro | 55.4% | 80.3% | 78.1% | 69.2% |
| LiveCodeBench (Pass@1) | 93.5% | 88.1% | 87.4% | 83.2% |
| Codeforces Elo | 3,206 | — | — | — |
| Terminal-Bench 2.1 | 83.9% | 88.0% | 85.1% | 82.7% |
Artificial Analysis data: Claude Fable 5 costs $3.48 per Strategy & Ops task (50 points); DeepSeek V4-Pro costs $0.03 (38 points) — a 116x cost gap with roughly a 12-point score difference.
DeepSeek V4 vs GPT-5.6 Sol and Claude Fable 5: Which Should You Pick?
| Dimension | DeepSeek V4-Pro | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| Open source / self-host | ✅ MIT | ❌ closed | ❌ closed |
| Context window | 1M tokens | not publicly disclosed | 1M tokens |
| Code capability | very strong (near Fable 5) | very strong | strongest (SWE-bench Pro 80.3%) |
| API output price (off-peak) | $0.87/M | ~$15/M | ~$50/M |
| Peak output price | $1.74/M | — | — |
| Agent capability | strong, multi-Agent orchestration | strong | strongest |
| Data privacy | ✅ on-prem deployable | ❌ | ❌ |
Tight budget / high call volume / private deployment → V4-Pro or V4-Flash; maximum code quality, cost no object → Claude Fable 5; complex algorithms + math reasoning → GPT-5.6 Sol / Ultra; massive log or document processing → V4-Flash (cache-hit input at just $0.0028/M).
Peak-Valley Pricing: How to Save Money and Six-Step API Migration Runbook (7/24 Deadline)
| Model | Billing item | Off-peak | Peak |
|---|---|---|---|
| V4-Pro | Input (cache hit) | ¥0.025 / $0.0035 / 1M | 2× rate |
| Input (cache miss) | ¥3.00 / $0.435 / 1M | ¥6.00 / $0.87 | |
| Output | ¥6.00 / $0.87 / 1M | ¥12.00 / $1.74 | |
| V4-Flash | Input (cache hit) | ¥0.02 / $0.0028 / 1M | 2× rate |
| Input (cache miss) | ¥1.00 / $0.14 / 1M | ¥2.00 / $0.28 | |
| Output | ¥2.00 / $0.28 / 1M | ¥4.00 / $0.56 |
Peak hours (Beijing time): weekdays 09:00–12:00 and 14:00–18:00. Even at peak, V4-Pro output ($1.74/M) remains 8.6× cheaper than Claude Opus 4.8 ($15/M).
client.chat.completions.create(
model="deepseek-v4-pro",
messages=[...],
extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)
Deadline warning: deepseek-chat → deepseek-v4-flash (non-thinking); deepseek-reasoner → deepseek-v4-flash (thinking mode) or deepseek-v4-pro. Legacy names stop working after July 24, 15:59 UTC.
Search the entire codebase: Find every reference to deepseek-chat and deepseek-reasoner, including environment variables and config files.
Map migration targets: Lightweight chat → deepseek-v4-flash; complex reasoning → deepseek-v4-pro with thinking parameters.
Validate in staging: Switch model names in pre-production, run regression suites, and confirm output quality holds.
Configure prompt cache: Place fixed system prompts at the top of message arrays to maximize cache hits (Flash cache-hit input at $0.0028/M).
Schedule batch workloads off-peak: Route document processing, code review, and other non-real-time jobs to after 18:00 or before 09:00 Beijing time.
Roll out to production: Complete full cutover before July 23, leaving a 24-hour rollback window. Watch for DeepSeek billing-change emails.
Is DeepSeek V4 GA Worth Upgrading? Citable Hard Data
CSA compression ratio: KV sequences pass through Softmax-gated pooling for 4× compression; V4-Pro uses top-1024 / V4-Flash top-512 sparse selection plus a 128-token sliding window.
HCA global attention: Tokens compress 128× before global dense attention; CSA and HCA alternate to capture long-range dependencies efficiently.
Cost anchor: V4-Flash stays below $0.04 per run across six index task categories; V4-Pro output at $0.87/M is the permanent post-June 75% price cut.
DeepSeek V4 GA is one of the most significant milestones in open-source AI for 2026. Its value is not about beating closed-source flagships on every benchmark — it is about delivering the strongest open performance at a fraction of the cost. For enterprises that cannot send data offshore, indie developers on tight budgets, and Agent engineers who need 1M-token context, V4-Pro is the most balanced choice available today.
If you plan to self-host V4-Flash locally or in the cloud for Agent routing and inference splitting, a consumer Mac with 16GB of RAM will not cut it — ds4 local inference documentation shows V4-Flash needs at least 96GB of unified memory. Pure API usage has no hardware barrier, but high-frequency Agent pipelines still demand host stability and 24/7 uptime: self-built VPS instances lack Metal acceleration, and swap thrashing causes inference timeouts; short-term trials rarely cover long-context stress tests. For a more stable production environment suited to iOS CI/CD and AI Agent automation, MESHLAUNCH Mac Mini cloud rental is usually the better path: dedicated Apple Silicon, on-demand upgrades to 64GB+ unified memory, 24/7 availability, and flexible daily/weekly/monthly billing.
Weekdays Beijing time 09:00–12:00 and 14:00–18:00 are peak hours with doubled rates. Run batch jobs after 18:00 to save. See the rental pricing page for cloud inference node options.
July 24, 2026 at 15:59 UTC (23:59 Beijing time). Migrate to deepseek-v4-flash or deepseek-v4-pro before then.
Pro fits complex reasoning and Agents; Flash fits lightweight routing with cache-hit input at $0.0028/M. Local deployment requirements are covered in the help center.
Fable 5 still leads on the hardest benchmarks, but V4-Pro's 80.6% on SWE-bench Verified is the open-source record — and output pricing runs about 1/17 of GPT-5.6 Sol.
Yes. V4 supports both OpenAI ChatCompletions and Anthropic Messages formats. Keep base_url at https://api.deepseek.com and update only the model parameter.