Overview
DeepSeek-V4-Flash Public Beta is the July 31, 2026 API update for DeepSeek's faster V4 model endpoint. The official update says API access is now in public beta for the formal DeepSeek-V4-Flash backend, while the request method stays the same: set the model name to deepseek-v4-flash to use the latest version. This release succeeds the April 2026 DeepSeek-V4 Preview for Flash API users, but it does not upgrade V4-Pro, DeepSeek APP, or DeepSeek WEB models yet.
The practical change is focused on agent reliability. DeepSeek says the model structure and size are unchanged from Preview, but the post-training process has been redone for the official Flash release. That makes this a low-friction upgrade for teams already using DeepSeek in coding agents, repo automation, tool-use workflows, and high-volume API workloads.
What's New
Formal Flash API Public Beta
DeepSeek-V4-Flash is now available as a formal public-beta API model. Existing API integrations do not need a new endpoint or request format; the official migration path is to keep using the deepseek-v4-flash model name. That matters for production teams because the upgrade can be rolled out behind normal model routing, eval, and rollback controls instead of a larger SDK migration.
DeepSeek also clarifies the scope of the release. Only the V4-Flash API has been upgraded in this update. V4-Pro API, the DeepSeek app, and DeepSeek web chat remain on their existing models, and DeepSeek says the V4-Pro formal release will follow later.
Stronger Agent Benchmarks
The July 31 update centers on Code Agent and tool-use performance. DeepSeek reports the following official benchmark scores for DeepSeek-V4-Flash under its evaluation harness:
| Benchmark | Official score |
|---|---|
| Terminal Bench 2.1 | 82.7 |
| NL2Repo | 54.2 |
| Cybergym | 76.7 |
| DeepSWE | 54.4 |
| Toolathlon verified | 70.3 |
| Agent Last Exam | 25.2 |
| Automation Bench (Public) | 25.1 |
| DSBench-FullStack | 68.7 |
| DSBench-Hard | 59.6 |
These numbers make the release most relevant to teams evaluating agentic coding, shell-based repair tasks, repository editing, automation workflows, and benchmark-driven model routing. DeepSeek notes that DSBench is an internal benchmark, so it should be treated as directional rather than independently reproducible public evidence.
Native Responses API Support
DeepSeek-V4-Flash supports the Responses API, which is important for Codex-style agent architectures that need structured tool calls, multi-step state, and traceable actions. The official update explicitly calls out targeted adaptation for Codex scenarios, making this release more practical for development agents than a generic chat-only model upgrade.
The pricing documentation lists additional integration surfaces for Flash, including JSON Output, Tool Calls, Anthropic API compatibility, prefix continuation beta, and Fill-in-the-Middle beta in non-thinking mode. V4-Pro does not yet list Responses API support in the same pricing table.
Same Architecture, New Post-Training
DeepSeek says the model structure and model size are the same as the Preview version, but the post-training process has been redone. For users, that means the release is not a new parameter-scale story or a new context-window story. It is an alignment, tool-use, and agent-behavior upgrade on the Flash endpoint.
This distinction matters for upgrade planning. If your current bottleneck is API cost, latency, tool reliability, or coding-agent success rate, V4-Flash Public Beta is the relevant update. If you were waiting for a stronger V4-Pro backend or a refreshed web/app model, this release is not that change.
Availability & Access
DeepSeek-V4-Flash Public Beta is available through the DeepSeek API. Use the model name deepseek-v4-flash to call the latest Flash version. The pricing page identifies the current Flash model version as DeepSeek-V4-Flash-0731.
Important availability notes:
- API upgraded: DeepSeek-V4-Flash API is the updated surface for this release.
- V4-Pro unchanged: V4-Pro API remains separate and has not received this July 31 public-beta upgrade.
- APP/WEB unchanged: DeepSeek app and web chat models are not changed by this release.
- Responses API supported: Flash supports Responses API, while the pricing table does not list the same support for V4-Pro.
- Public beta: Treat the release as production-evaluable but still subject to public-beta behavior, docs updates, and possible pricing-policy notices.
Pricing & Plans
The DeepSeek pricing page lists deepseek-v4-flash as a token-based API model with 1M context, 384K maximum output, and a concurrency value of 2,500. Pricing is denominated in CNY per million tokens:
| Model | Cache-hit input | Cache-miss input | Output | Context | Max output |
|---|---|---|---|---|---|
| deepseek-v4-flash | CNY 0.02/M | CNY 1/M | CNY 2/M | 1M tokens | 384K tokens |
| deepseek-v4-pro | CNY 0.025/M | CNY 3/M | CNY 6/M | 1M tokens | 384K tokens |
Flash remains the lower-cost V4 API choice, especially for cached-input workflows and high-volume coding agents. DeepSeek also states that peak and off-peak pricing is planned: peak periods are expected to charge 2x the listed price during Beijing time 09:00-12:00 and 14:00-18:00, but the pricing page says formal notice will come later.
Best For
- Coding-agent teams already calling
deepseek-v4-flashwho want better tool-use performance without rewriting API requests. - Developers building Codex-style agents around Responses API, tool calls, shell tasks, and repo edits.
- High-volume API workloads where Flash's CNY pricing and 2,500 listed concurrency matter.
- Evaluation teams comparing agent benchmarks such as Terminal Bench, NL2Repo, Cybergym, DeepSWE, and Toolathlon.
- Users who need 1M context and large output limits but do not need V4-Pro pricing or future Pro capability.
FAQ
What changed in DeepSeek-V4-Flash on July 31, 2026?
DeepSeek upgraded the deepseek-v4-flash API to the formal public-beta Flash version. The request method and model name stay the same, but DeepSeek says the post-training process has been redone and agent benchmark performance is stronger.
Is DeepSeek-V4-Pro upgraded too?
No. DeepSeek says this update upgrades only the V4-Flash API. V4-Pro API, DeepSeek APP, and DeepSeek WEB models remain unchanged, with the formal V4-Pro release expected later.
What model name should I use?
Use deepseek-v4-flash. The official update says setting the model name to deepseek-v4-flash is enough to call the latest Flash API version.
How much does DeepSeek-V4-Flash cost?
The official pricing page lists CNY 0.02 per million cache-hit input tokens, CNY 1 per million cache-miss input tokens, and CNY 2 per million output tokens for deepseek-v4-flash.
Does DeepSeek-V4-Flash support the Responses API?
Yes. DeepSeek lists Responses API support for Flash, and the July 31 update calls out targeted adaptation for Codex-style agent use cases.
