DeepSeek-V4-Flash Official API Launch! Agent Capability Soars 6x, Price Only 1/90 of Claude

Chinese AI startup DeepSeek has officially launched the DeepSeek-V4-Flash API. According to the company, the new version significantly enhances AI agent capabilities without altering the original model architecture or parameter scale—achieving this solely through re-optimized post-training. Notably, its performance on the DeepSWE benchmark surged from 7.3 points in the V4-Pro Preview three months ago to 54.4 points—an increase of over 6 times. Meanwhile, the API maintains an extremely low pricing model, with input token costs as low as 0.02 RMB per million tokens on cache hit, drawing intense interest from the developer community.

DeepSeek states that the V4-Flash official version adopts a Mixture of Experts (MoE) architecture, with a total of 284 billion parameters and approximately 13 billion active parameters. It supports a 1-million-token context length and allows switching between thinking and non-thinking modes.

Among nine agent-related benchmark tests released by the company, the most dramatic improvement was seen in DeepSWE, where scores jumped from 7.3 to 54.4. In Terminal Bench 2.1, V4-Flash scored 82.7 points—surpassing V4-Pro Preview’s 72.1 and GLM-5.2’s 81.0, and trailing only slightly behind Claude Opus 4.8’s 85 points.

In the Intelligence Index test published by Artificial Analysis, V4-Flash-0731 achieved a score of 50 in its highest reasoning mode, far exceeding the median score of 17 among comparable models, indicating that its overall agent capabilities have reached a leading level.

The most distinctive aspect of this release is that DeepSeek kept the model backbone unchanged. The company emphasized that V4-Flash-0731 retains the same architecture, size, and parameter scale as the preview version—the only difference being the re-execution of the post-training process. This has sparked renewed discussion about whether the future of large language model competition will shift from "scaling parameters" to "optimizing training efficiency."

DeepSeek also introduced its self-developed agent evaluation framework, DeepSeek Harness, enabling developers to validate model performance in multi-step tasks, autonomous planning, and tool-calling scenarios.

On pricing, DeepSeek continues its low-cost strategy. Input tokens cost 0.02 RMB per million on cache hit, 1 RMB on miss, and output tokens cost 2 RMB. In USD, the output cost is approximately $0.28—just 1/90th of Anthropic’s Claude Opus 4.8, which charges around $25 per million output tokens.

Many developers have shared real-world test results. One user reported running the official API alongside Claude Code continuously for nearly four hours, consuming about 100 million tokens, with a 99% cache hit rate and total cost under 1 RMB, while praising the high code quality and significantly reduced rework. Another developer noted the model could write code and fix over ten bugs in one go, costing only about $0.10.

Engineers working on Qt/C++ and STM embedded development commented that V4-Flash-0731’s overall capability is close to V4-Pro Preview but at a much lower price, offering exceptional cost-effectiveness.

However, V4-Flash is currently only available via public API testing—web and app versions remain unupdated, so general users cannot yet access it directly. Some developers have pointed out areas for improvement, such as instruction following accuracy, language control in agent mode, and embedded code quality. For example, inference language is currently fixed in English, and System Prompts do not function fully. On GitHub, users have requested more language control parameters.

From an engineering perspective, V4-Flash now officially supports the OpenAI Responses API, enabling direct integration with tools like Codex CLI, ChatGPT desktop, and VS Code Codex extensions—reducing developer onboarding costs. DeepSeek noted that Responses API support is currently limited to V4-Flash, with V4-Pro expected to gain support by early August 2026.

Regarding long-context computational efficiency, DeepSeek claims that in 1-million-token scenarios, the V4 series reduces per-token inference computation to 27% of V3.2’s level, with KV Cache usage dropping to about 10%. This greatly lowers GPU compute and memory demands for long-chain tasks.

Notably, on the same day DeepSeek announced the V4-Flash official release, OpenAI also revealed an 80% price cut for its GPT-5.6 Luna API. Market analysts believe global AI model price competition is rapidly intensifying. DeepSeek’s achievement of a major agent capability leap through post-training—without increasing model size—signals that the LLM race is increasingly shifting from "parameter scaling" to "training efficiency optimization."

Comparison of DeepSeek-V4-Flash and V4-Pro API Specifications:

| Item | DeepSeek-V4-Flash | DeepSeek-V4-Pro | |---|---|---| | OpenAI Format Base URL | https://api.deepseek.com | https://api.deepseek.com | | Anthropic Format Base URL | https://api.deepseek.com/anthropic | https://api.deepseek.com/anthropic | | Model Version | DeepSeek-V4-Flash-0731 | DeepSeek-V4-Pro | | Thinking Mode | Supports toggle between thinking/non-thinking | Supports toggle between thinking/non-thinking | | Context Length | 1 million tokens | 1 million tokens | | Max Output | 384K tokens | 384K tokens | | JSON Output | Supported | Supported | | Tool Calls | Supported | Supported | | Responses API | Supported | Not supported (expected early August 2026) | | Anthropic API | Supported | Supported | | Dialogue Prefix Continuation (Beta) | Supported | Supported | | FIM Completion (Beta) | Supported only in non-thinking mode | Supported only in thinking mode | | Input per Million Tokens (Cache Hit) | 0.02 RMB | 0.025 RMB | | Input per Million Tokens (Cache Miss) | 1 RMB | 3 RMB | | Output per Million Tokens | 2 RMB | 6 RMB | | Max Concurrent Requests | 2,500 | 500 |

Supplementary Notes:

| Item | Content | |---|---| | Responses API | Currently only supports DeepSeek-V4-Flash; V4-Pro support expected by early August 2026 | | Peak Pricing | Peak rates apply during high-demand periods; peak price is double the standard rate | | Peak Hours | Beijing time: 9:00–12:00, 14:00–18:00 daily |