Update — July 1, 2026: DeepSeek's official announcement confirms mid-July 2026 for V4 official — not a new model name, but the graduation from preview with peak-hour pricing and stated 功能优化和性能提升 (feature optimization and performance improvements). Live pricing docs · V4 preview migration guide. Last updated: July 1, 2026.
If you have been calling deepseek-v4-pro or deepseek-v4-flash since April 2026, you were on preview — not the official release. Teortaxes (@teortaxesTex) put it bluntly on June 29: "Yeah yeah you might think we had V4 for over 2 months already, but no, that was 'preview of V4.'"
DeepSeek now expects heavy demand at official launch and is introducing peak-hour pricing: 2× baseline during defined Beijing business windows. Off-peak rates stay the same as today's preview pricing.
TL;DR — DeepSeek V4 Official vs Preview
| Item | Detail |
|---|---|
| Launch window | Mid-July 2026 (7月中旬) — official V4 |
| What you had since April | V4 Preview — same model IDs, not final release |
| Peak hours (Beijing) | 9:00–12:00 and 14:00–18:00 daily |
| Peak multiplier | 2× off-peak baseline on all listed token prices |
| Off-peak baseline | Unchanged from current preview pricing |
| Notice | Email 24 hours before pricing takes effect |
| Opt-out | Stop service + apply for refund of remaining balance |
| Stated improvements | 功能优化 + 性能提升 (feature optimization + performance lift) |
| Legacy IDs | deepseek-chat / deepseek-reasoner still retire July 24, 2026 |
The Official Announcement — What DeepSeek Said
DeepSeek's Chinese notice (June 29, 2026) states:
The official version of DeepSeek V4 is planned to launch in mid-July. This version update will bring more feature optimizations and performance improvements. To ensure stable service quality and reasonable resource allocation, the API will introduce a peak and off-peak pricing mechanism.
Peak hours defined
| Window | Beijing time |
|---|---|
| Morning peak | 09:00 – 12:00 |
| Afternoon peak | 14:00 – 18:00 |
| Off-peak | All other hours |
During peak hours, listed prices double. During off-peak, prices match the current baseline.
User protections
- 24-hour email notice before pricing adjustment takes effect
- Refund path: users who disagree may stop using the service and apply for a refund of remaining balance
Full Pricing Tables — CNY (per million tokens)
Baseline = off-peak. Peak = 2× baseline.
deepseek-v4-pro
| 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 |
deepseek-v4-flash
| 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 |
USD baseline (off-peak) — current preview rates
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 |
Peak USD: multiply off-peak by 2× during Beijing peak windows.
Even doubled, Teortaxes and replies note V4 remains "dirt cheap" vs US frontier APIs — especially Flash for high-volume agent loops. See our V4-Pro economics deep-dive for agent workload math.
Preview vs Official — What Actually Changes?
Same model IDs, new commercial frame
Since April 24, 2026 preview launch, developers have used:
deepseek-v4-pro— 1.6T MoE class, 49B active (vendor-reported)deepseek-v4-flash— 284B / 13B active, economical tier- 1M context, 384K max output, thinking + non-thinking modes
- OpenAI + Anthropic-compatible API surfaces
Official launch does not rename these IDs. It adds:
- Peak/off-peak billing — demand management
- Stated 功能优化 (feature/UX polish) and 性能提升 (performance lift)
Reading the Chinese — polish or intelligence?
Teortaxes read the announcement phrasing as "no new features, just continued infra optimization and model polish."
That is one fair reading of 功能优化 (functional optimization — smoother API, stability, caching). But 性能提升 is ambiguous in Chinese AI product language:
| Term | Narrow read | Broad read |
|---|---|---|
| 功能优化 | API polish, bug fixes, UX | Refined tool-calling, JSON modes |
| 性能提升 | Faster TPS, lower latency | Higher benchmark accuracy, better reasoning |
| 版本更新 | Software patch | New model deployment |
Community replies like @xhyctf asked directly: "performance improvement should be pretty significant, right?" DeepSeek has not attached new benchmark tables to the pricing notice — treat performance claims as directional until weights/docs update.
What preview already gave you
| Capability | Preview (Apr 2026 – mid-Jul 2026) |
|---|---|
| 1M context | Default on official API |
| Open weights | Hugging Face collection |
| Agent integrations | Claude Code, OpenClaw, OpenCode (per DeepSeek) |
| Legacy mapping | deepseek-chat → v4-flash non-thinking |
Official mid-July is the production billing and stability milestone — not necessarily a wholly new architecture drop.
Timezone Math — Who Gets Cheap Hours?
Peak windows are Beijing time. That creates predictable pain:
| Region | Local overlap with Beijing peak |
|---|---|
| China | Business hours — peak |
| US East (EDT) | Roughly 21:00–02:00 and 02:00–06:00 — mixed |
| US West (PDT) | Evening/night — often off-peak for US devs |
| Europe (CEST) | Morning/afternoon — partial peak overlap |
@Shoier__ noted the squeeze: "During the day in China it's peak hour. During the day in 'Murica it's also peak hour. So it leaves not so many hours where it is cheap."
@GeorgeBisbas countered optimistically: "Europe can fully have non-peak at their working hours."
Practical takeaway: US West Coast and late-shift EU teams can batch heavy agent jobs off-peak. China-native production traffic pays peak by design — DeepSeek's revenue model targets domestic business-hour demand.
Teortaxes Thread — Context From a DeepSeek Watcher
Teortaxes' June 29 post (153K+ views) framed the move in DeepSeek pricing history:
- Previously: higher R1 vs V3 pricing, off-peak discounts, permanent discounts, flat rate cuts
- Peak pricing is structurally inverse off-peak discount — same lever, opposite direction
- "In any case I approve, they need revenue" — aligns with Chinese AI bubble sustainability debates
DeepSeek undercut Western APIs for two years. Peak pricing is the first widespread surge-pricing mechanism on V4 — still cheap at 2×, but no longer flat-rate infinite scale at preview prices during CN business hours.
Model Specs Unchanged — Quick Reference
From DeepSeek's Models & Pricing page (current preview docs):
| Spec | deepseek-v4-flash | deepseek-v4-pro |
|---|---|---|
| Context | 1M tokens | 1M tokens |
| Max output | 384K | 384K |
| Thinking modes | Both (default thinking) | Both (default thinking) |
| Tool calls / JSON | Yes | Yes |
| FIM completion (beta) | Non-thinking only | Non-thinking only |
| Concurrency limit | 2,500 | 500 |
Legacy retirement: deepseek-chat and deepseek-reasoner → July 24, 2026 (migration guide).
What Builders Should Do Before Mid-July
1. Budget with peak multipliers
Model agent workloads that run during Beijing 9–12 / 14–18 at 2× token cost. Cache hits stay cheap even at peak (e.g. Pro cache hit peak = ¥0.05/M).
2. Shift batch jobs off-peak where possible
Embeddings sweeps, eval runs, dataset generation — schedule outside peak if your timezone allows.
3. Watch your email
DeepSeek promises 24-hour notice — do not get surprised on billing day.
4. Decide on refund vs continue
Disagree with surge pricing? Withdraw balance per official policy before running production peaks.
5. Keep preview evals running
Official may improve prompt following (community complaint: "super smart kid with ADHD"). Re-benchmark on your tasks when mid-July drops — not X speculation.
6. Plan legacy ID migration
July 24 retirement still looms independent of mid-July official launch.
DeepSeek V4 in the Broader June 2026 Landscape
Mid-July V4 official lands in a crowded open-weight market:
- GLM-5.2 — BridgeBench reasoning leader post-Fable ban
- Kimi K2.7-Code — agentic coding, Modified MIT
- LongCat-2.0 — 1.6T MoE, weights pending
- Fable 5 still offline — US export control pushes international devs to DeepSeek stack
DeepSeek's move is monetize demand at peak while keeping off-peak floor that already disrupted Western pricing in Q1 2026 (pricing disruption post).
The Honest Answer
Is DeepSeek V4 "new" in mid-July?
Commercially yes, technically incremental. Preview → official with peak pricing and stated optimizations. Do not expect a surprise V5 name change.
Will prices go up?
Yes, half the day. Peak = 2×. Off-peak = same as today.
Will the model get smarter?
Maybe. Announcement language allows it; no new public benchmarks shipped with the pricing notice. Wait for mid-July docs and re-run your evals.
Is it still worth it?
For most agent builders, even peak Flash pricing undercuts US frontier APIs by an order of magnitude. Pro at peak is still a fraction of Claude/GPT tier pricing — with 1M context and open weights.
Update — July 31, 2026: DeepSeek shipped DeepSeek-V4-Flash-0731 — an agent-capability upgrade to the Flash API tier with native Responses API support and full Codex compatibility, independently benchmarked by Artificial Analysis at ~2-3x cheaper than OpenAI Luna for similar intelligence. V4-Pro remains unchanged.
Update — August 6, 2026: DeepSeek warned of a coming "significant" API price increase across its products, with no exact rates disclosed yet. See DeepSeek's price increase warning for what's known so far.
Related Reading
- DeepSeek's API price increase warning (Aug 2026)
- DeepSeek-V4-Flash-0731: Codex Support and $0.14/$0.28 Pricing
- DeepSeek V4 Preview — API Migration
- DeepSeek V4-Pro Benchmarks & Agent Economics
- DeepSeek V4 Pricing Disruption
- China AI Playbook — Free Models, Cheap Compute
- GLM-5.2 vs Fable 5
Pricing tables reflect DeepSeek's June 29, 2026 announcement and api-docs.deepseek.com Models & Pricing page. Peak hours, dates, and rates may change — verify official docs before production budgeting.