DeepSeek V4 API: Specs, Pricing, and What the V4-Flash-0731 Release Means for Developers

On July 31, 2026, DeepSeek moved V4 Flash out of preview with an official release called DeepSeek-V4-Flash-0731, served through the API under the model ID deepseek-v4-flash. It is the first officially released V4 model — deepseek-v4-pro remains in preview — and brings a 1M-token context window, 384K max output, and per-token pricing that undercuts most frontier competitors.

This week's release is a pricing and efficiency event as much as a capability update. Here is what changed, what the DeepSeek V4 API costs, and what developers should watch next.

What the DeepSeek V4 API ships now

Both V4 models are live on the official DeepSeek API with OpenAI-compatible and Anthropic-compatible endpoints:

Capability	deepseek-v4-flash	deepseek-v4-pro
Build version	DeepSeek-V4-Flash-0731 (official release)	DeepSeek-V4-Pro (preview)
Context window	1M tokens	1M tokens
Max output	384K tokens	384K tokens
Thinking mode	Non-thinking and thinking (default)	Non-thinking and thinking (default)
JSON output / tool calls	Yes	Yes
Responses API	Yes	Coming early August 2026
Anthropic API format	Yes	Yes
FIM completion (beta)	Non-thinking only	Non-thinking only

The most notable spec is the 1M-token context window with a 384K max output — enough to process entire codebases, long agent logs, or book-length documents in a single request. Thinking mode defaults to on, but can be switched off per request for latency-sensitive workloads. The official V4-Flash API also natively supports the Responses API format and is specifically adapted for Codex, according to DeepSeek's API changelog.

V4 Flash is an efficiency-optimized Mixture-of-Experts model with 284B total parameters and 13B activated per token, according to its OpenRouter model page. DeepSeek describes the architecture as combining hybrid attention for efficient long-context processing with fast inference and high throughput, making it a fit for coding assistants, chat systems, and agent workflows where responsiveness matters.

DeepSeek V4 API pricing: the numbers that matter

DeepSeek publishes V4 pricing in units of per 1M tokens on its official pricing page:

Model	Input (cache miss)	Input (cache hit)	Output
deepseek-v4-flash	$0.14 / 1M	$0.0028 / 1M	$0.28 / 1M
deepseek-v4-pro	$0.435 / 1M	$0.003625 / 1M	$0.87 / 1M

Three things stand out:

• Flash is priced for scale. At $0.14 in / $0.28 out per million tokens, V4 Flash sits far below most frontier models, and the 2500-request concurrency limit on the official API supports high-throughput workloads (Pro is capped at 500).
• Prompt caching is the real discount. Cache-hit input drops to $0.0028 per 1M tokens on Flash — 50x cheaper than a cache miss. Agent loops and RAG pipelines that resend the same system prompt and context see the biggest savings.
• Peak-hour pricing is coming. DeepSeek announced a peak/off-peak policy: during peak hours (9:00–12:00 and 14:00–18:00 Beijing Time, UTC+8, daily) all billing items will be 2x the regular price. The effective date has not been announced yet, but teams with flexible scheduling should plan for it.

Background: from preview to production

DeepSeek V4 was first introduced as a preview release on April 24, 2026, when the company announced the new model generation with a smaller parameter size, faster response times, and lower API pricing. The preview period also included a transition: the legacy model aliases deepseek-chat and deepseek-reasoner were discontinued on July 24, 2026, with calls migrated to the V4 models.

The July 31 production release of V4-Flash-0731 ends that preview chapter. DeepSeek's API changelog and third-party coverage, including this breakdown of the 0731 build, describe it as the official Flash release — same model ID, re-post-trained weights, and substantially stronger agent benchmarks — with the official V4-Flash API itself now in public beta. The weights are open under an MIT license on Hugging Face, continuing DeepSeek's open-source pattern.

Impact: what it means for developers and the market

The V4-Flash-0731 launch arrives in the middle of an aggressive AI pricing cycle. Analysts have framed the $0.28 output floor on Flash as a deliberate move in the pricing war around agentic workloads — long-horizon agents that burn large amounts of tokens need cheap per-token economics to be commercially viable (Yahoo Finance).

For developers, the practical takeaways are:

• Agent and coding workloads get cheaper. The combination of 1M context, thinking mode, tool calls, and $0.28/1M output makes Flash a natural default for coding assistants, automated QA, and multi-step agents — the exact workloads that dominated frontier API bills in 2026.
• Long-context RAG is finally affordable. With cache-hit input at $0.0028/1M, repeatedly sending a large knowledge base into the prompt becomes cost-effective instead of prohibitive.
• Multi-provider availability is a safety net. Beyond the official API, V4 Flash is served by over 20 providers through OpenRouter at a discounted blended rate ($0.0882 in / $0.1764 out per 1M at the time of writing), which gives teams redundancy and pricing leverage.

The main uncertainty is operational, not technical: peak-hour pricing at 2x could materially change monthly bills for teams with China-daytime traffic patterns, so the effective date is worth tracking closely.

What to watch next

Three items are on the near-term roadmap:

1. Responses API support for deepseek-v4-pro — DeepSeek says it will add Responses API support for Pro in early August 2026; it currently supports Flash only.
2. The peak/off-peak pricing effective date — officially announced but not yet scheduled; the current announcement is on the pricing docs page.
3. V4 Pro production status — while Flash has reached its official release, Pro remains in preview, and DeepSeek says the official V4-Pro release will follow soon. Its $0.87/1M output price already undercuts most top-tier rivals.

For teams evaluating the switch, the DeepSeek API quick start documents the OpenAI-compatible base URL (https://api.deepseek.com), thinking-mode switching, and the Anthropic-format endpoint — most existing SDKs work by swapping the model name.

The DeepSeek V4 API has effectively reset expectations for what a frontier-class model should cost. With Flash now in production, cheap 1M-token context, and cache pricing that makes agents viable, the next few months will show whether competitors match the price or the performance.

This article is part of our AI industry news coverage. For AI video generation tools and workflows, visit wan27.org.