DeepSeek V4 Goes Official: Peak-Hour Pricing and the July 24 Migration Deadline

DeepSeek V4 shipped its general-availability release on July 19, 2026, and the launch is less about benchmark crowns than about a structural change in how AI compute is sold. Alongside the model upgrade, DeepSeek introduced peak-hour pricing and set a hard July 24 migration deadline for legacy API endpoints. For production teams, the message is clear: cheap tokens are still available, but only if you schedule around Beijing business hours and stop calling endpoints that will vanish in five days.

What Changed in V4 General Availability

The GA release follows the V4 Preview that has been live since April 24, 2026. DeepSeek offers two API models: deepseek-v4-pro at 1.6 trillion total parameters with 49 billion active, and deepseek-v4-flash at 284 billion total parameters with 13 billion active. Both come with a 1-million-token context window, thinking and non-thinking modes, and OpenAI- or Anthropic-format endpoints.

DeepSeek claims stronger agent-based task execution, mathematical reasoning, and generation quality in the GA build. Independent early tests show 80.6% on SWE-bench and a 3,206 Codeforces rating, though the broader community is waiting for Chatbot Arena and Artificial Analysis results to confirm those figures. The model weights are open under an MIT license, which has already sparked a wave of community fine-tunes and quantized deployments.

Peak-Hour Pricing Is the Real Headline

The most consequential change is not the model; it is the meter. DeepSeek now charges 2x during peak hours, defined as 9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m. daily. Off-peak rates remain unchanged. For V4 Pro, that means off-peak output tokens cost roughly $0.82 per million, while peak output jumps to about $1.64 per million. Cache-hit input prices stay low, but the gap between cache hit and cache miss can reach 120x, making prompt caching a make-or-break optimization.

This is the first time a major frontier lab has introduced time-of-day pricing at scale. It signals that inference capacity, not model training, is now the binding constraint. DeepSeek can train a better model, but serving unlimited concurrent requests during Chinese business hours without latency degradation is a harder problem. The pricing structure is how the lab manages demand.

The July 24 Migration Deadline

Buried in the release notes is a hard cutoff: after July 24, 2026, at 15:59 UTC, the legacy endpoints deepseek-chat and deepseek-reasoner will stop responding. During the Preview phase, deepseek-reasoner mapped to V4 Flash, not V4 Pro. Teams that need heavier reasoning must now explicitly call deepseek-v4-pro.

The migration is technically simple, update the model field in your API call and keep the base URL. The operational risk is not trivial. Any production system still pointing at the old endpoint names will fail silently after the deadline. Engineering teams should audit their code this week, not next month.

Where V4 Fits in a Crowded July

July 2026 has been unusually dense with releases. Kimi K3 shipped on July 14. GPT-5.6 Sol arrived mid-month. Fable 5 and GLM-5.2 both launched in early July. V4's angle is straightforward: open-weight pricing with near-closed-source performance. The trade-offs are real: long-context retrieval still trails Opus 4.8, specialized software engineering remains Fable 5's territory, and complex agent tasks need more iteration rounds than Claude.

For teams deciding where to deploy, the calculus is now hourly as well as architectural. Interactive traffic during Chinese peak hours absorbs the higher rate. Offline evaluation runs, bulk document processing, and synthetic-data generation should shift to off-peak windows. Teams outside China may benefit unintentionally if peak windows map to Beijing time, but DeepSeek has not clarified timezone rules, so cost projections should include a safety margin.

A separate debate has shadowed the rollout. Independent researcher @synthwavedd published an investigation claiming that under specific conditions, DeepSeek V4 outputs are virtually identical to Claude Fable 5, suggesting possible distillation. DeepSeek has not responded publicly, and the claims remain contested. The practical takeaway is the same as always: run your own benchmarks. You can find a curated list of AI models and inference hosts, including alternatives to DeepSeek, in the AI tools directory at aifreetool.site.

Key Takeaways and My Take

• Schedule your traffic. Peak-hour pricing makes prompt timing a first-class cost lever. Move batch work to off-peak hours.
• Migrate before July 24. The legacy endpoints will stop working. Audit integrations now and explicitly choose Pro or Flash.
• Optimize caching. The 120x gap between cache hit and cache miss means prompt-template reuse is now a direct budget win.
• Benchmark on your own data. Published SWE-bench and Codeforces scores are useful, but workload-specific tests matter more.

My take: DeepSeek V4 GA is a milestone for open-weight AI, but the pricing change is the bigger story. It proves that even a lab famous for undercutting rivals now faces capacity constraints. That is good news for the open ecosystem, because it means demand is real. It is also a warning: the era of flat, unlimited cheap tokens is ending. Smart teams will treat inference scheduling as seriously as they treat model selection.

Frequently Asked Questions