Frontier Models

DeepSeek-V4-Flash API Enters Public Beta With 1M Token Context

The July 31 release upgrades agent performance on multiple benchmarks while maintaining the same model architecture as the preview and introducing native Responses API support at competitive per-token rates.

By Marcus Vance July 31, 2026 4 MIN READ

In a modern open-plan technology development office with floor-to-ceiling windows overlooking an urban skyline at dusk, an anonymous software engineer wearing a plain dark hoodie and jeans sits at a clean wooden desk facing away from the camera. The engineer is focused on a laptop displaying abstract graphical interfaces without any readable content, surrounded by multiple external hard drives and network cables neatly organized on the desk surface. Behind the engineer, tall server racks filled with blinking indicator lights and cooling fans line the wall, representing backend infrastructure for large-scale AI model deployment. The room features neutral gray walls, ergonomic office chairs, potted plants, and scattered technical reference books on shelves. On the desk are additional items including a wireless mouse, a notebook with blank pages, a coffee mug, and USB storage devices. The overall environment conveys a professional workspace dedicated to testing and integrating new API capabilities for agent-based systems, with emphasis on expansive data handling capacity through the visible scale of connected hardware and the engineer's concentrated posture during a late workday session. Subtle reflections from overhead LED panels illuminate the scene evenly, highlighting the texture of the wooden desk grain, the matte finish of the server enclosures, the fabric of the hoodie, and the precise arrangement of cables running from the laptop to peripheral devices. The background includes additional workstations with similar setups, empty meeting rooms visible through glass partitions, and a distant view of city lights suggesting ongoing global connectivity. This composition centers on the tangible elements of AI infrastructure rollout, including the physical hardware supporting high-capacity context processing and response generation features introduced in recent software updates. The engineer’s hands rest on the keyboard in a natural typing position, while the server racks extend into the depth of the frame, underscoring the robust, scalable nature of the technology platform being prepared for broader developer access and performance evaluation on standard benchmarks. — Illustration: AI Intel Report

DeepSeek-V4-Flash is a Mixture-of-Experts language model with 284 billion total parameters and 13 billion active parameters released in public beta through the DeepSeek API on July 31, 2026.

The DeepSeek-V4-Flash API entered public beta on July 31, 2026, via the DeepSeek API platform. The model name deepseek-v4-flash was updated to DeepSeek-V4-Flash-0731 while preserving the identical architecture and parameter count from the preview version. Only re-post-training was applied to improve agent performance. This release focuses exclusively on the Flash API endpoint.

A 1 million token context window allows processing of entire large codebases or lengthy documents within a single interaction. The maximum output of 384,000 tokens supports generation of extended responses such as full application code or comprehensive analysis reports. These specifications target agent workflows that require sustained context across multiple steps.

What technical specifications distinguish DeepSeek-V4-Flash from prior versions?

The model provides native support for the Responses API format along with tool calling and JSON output modes. It is also adapted for Codex, enabling direct use in coding agent environments. These features reduce the need for custom wrappers when integrating the model into existing agent frameworks.

The Mixture-of-Experts design activates only 13 billion parameters out of the total 284 billion during inference. This selective activation maintains computational efficiency while delivering high performance on specialized tasks. The approach contributes directly to the low per-token pricing offered in the public beta.

How do the benchmark results demonstrate enhanced agent capabilities?

DeepSeek-V4-Flash recorded a score of 82.7 on Terminal Bench 2.1. This benchmark measures performance in terminal command and agent-driven task execution. The result indicates reliable handling of multi-step command sequences.

On the Toolathlon Verified benchmark the model achieved 70.3. This score reflects accuracy in verified tool invocation scenarios. Strong results here point to improved reliability when agents must select and execute external tools.

The DSBench-FullStack benchmark produced a score of 68.7. This evaluation covers full-stack development workflows. The outcome shows capability across integrated development tasks that combine frontend, backend, and deployment steps.

Benchmark performance of DeepSeek-V4-Flash reported in the official change log
Benchmark	Score	Focus Area
Terminal Bench 2.1	82.7	Terminal and command agent tasks
Toolathlon Verified	70.3	Verified tool use accuracy
DSBench-FullStack	68.7	Full-stack development workflows

What key features receive native support in the Responses API?

Responses API format compatibility
Tool calling functions
JSON structured output
Codex adaptation for coding agents
1M token context handling

What pricing applies to the DeepSeek-V4-Flash API?

Input tokens are priced at $0.14 per million on cache miss. Output tokens are priced at $0.28 per million. The Responses API currently supports only the deepseek-v4-flash model under this pricing schedule.

How does the release influence the frontier models market?

The combination of large context length, high output capacity, and low pricing intensifies competition on efficiency metrics. Developers building long-running agent systems gain access to extended context without proportional cost increases. The native tool calling and Responses API support lowers integration barriers for existing agent platforms.

Stakeholders in enterprise AI deployments may evaluate the model for tasks that previously required multiple shorter calls or external memory systems. The 13 billion active parameters deliver performance levels that challenge denser models at higher price points.

What reactions accompanied the public beta announcement?

The official statement from DeepSeek highlighted upgrades to agent capabilities and benchmark improvements over the V4-Pro-Preview. Emphasis was placed on the native Responses API support and Codex adaptation as immediate integration benefits.

🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta! 🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. ... 🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!DeepSeek, Official account (@deepseek_ai)

What developments are anticipated next from DeepSeek?

The official release of DeepSeek-V4-Pro will follow soon after the current Flash beta. The present update affects only the Flash API variant while Pro endpoints and web models stay unchanged. Further post-training refinements on the same architecture are expected to continue.

Developers monitoring the frontier models space will track whether the Pro version introduces additional parameter scaling or new feature sets while retaining the pricing and context advantages demonstrated in the Flash release.

Frequently asked

When did the DeepSeek-V4-Flash API enter public beta?

The DeepSeek-V4-Flash Official API entered public beta on July 31, 2026, via the DeepSeek API platform.

Sources

DeepSeek — The official release of the DeepSeek-V4-Flash API is now in public beta with benchmark results Terminal Bench 2.1: 82.7, Toolathlon verified: 70.3, DSBench-FullStack: 68.7, and native Responses API support.
DeepSeek — CONTEXT LENGTH 1M, MAX OUTPUT MAXIMUM: 384K, PRICING 1M INPUT TOKENS (CACHE MISS) $0.14, 1M OUTPUT TOKENS $0.28, with Responses API and Tool Calls support.
X — DeepSeek announced the public beta release and benchmark improvements over V4-Pro-Preview along with Responses API and Codex support.