DeepSeek V4 Flash Arrives, with Pro on the Horizon

This is a text edition. Images, polls and embedded content from the original are not yet included; see the original source below.

Rainy-Day Reflections · AI Observation DeepSeek V4 Flash Official Release Arrives Much stronger agent capabilities, plus support for the Responses API NEW · Released July 31, 2026 Quick report · August 1, 2026

DeepSeek V4 Flash Official Release Arrives: Stronger Agents and Responses API Support

In one sentence: on the last day of July, DeepSeek quietly announced a public beta of the official V4-Flash API. Its main selling points are substantially improved agent capabilities and new support for the Responses API format; model weights were also released under the MIT license, while prices remain unchanged.

Key information Version: DeepSeek-V4-Flash-0731 · API name unchanged Status: public beta of the official-release API Open source: weights on Hugging Face · MIT license Context: 1M tokens · maximum output 384K Architecture: 284B-parameter mixture of experts · 13B active parameters

01 NEWS What happened: a release on the last day of July

On July 31, 2026, the changelog in DeepSeek's official API documentation published 'DeepSeek-V4-Flash Update,' announcing that the official-release API had entered public beta, with model version DeepSeek-V4-Flash-0731. The homepage simultaneously displayed a banner: 'The official V4-Flash API is now in public beta, with greatly improved agent capabilities; V4-Pro is unchanged.'

Only the Flash API was upgraded this time—V4-Pro and the models in the app and on the website were unchanged. The official V4-Pro release was said to be coming 'as soon as possible,' with Responses API and Codex support expected in early August. According to the official explanation, its architecture and scale are the same as Preview's; 'only the post-training was redone.'

02 BENCH Where the gains are: agent benchmarks soar

The official announcement made agent capability its main selling point. The tests used the minimal mode of the soon-to-be-open-sourced DeepSeek Harness (max effort). V4-Flash-0731 substantially outperformed the much larger V4-Pro Preview on several agent benchmarks:

Most strikingly, its DeepSWE score rose from 7.3 for Flash Preview to 54.4, nearly 7.5 times as high. Third-party Artificial Analysis gave it an Intelligence Index score of 50, third among 101 models, comparable with Gemini 3.6 Flash, ten points above Preview and only one point below GPT-5.6 Luna and GLM-5.2.

This still deserves a sober reading. Results vary by benchmark: someone on Hacker News noted that GPT-5.6 Luna still scored 67% on DeepSWE, above Flash's 54.4%, so this is not a clean sweep. The official figures mainly cover agent and coding benchmarks; general results for MMLU, AIME, LiveBench, and the like had not yet been released.

03 API The key phrase: Responses API support

For developers, the most informative part of the update is the official documentation's explicit addition of 'support for the Responses API format.' The endpoint remains api.deepseek.com; for now, only deepseek-v4-flash is available, with V4-Pro planned for early August.

WHAT IS THE RESPONSES API? The Responses API is a new-generation 'unified interface' first introduced by OpenAI, intended gradually to replace the older Chat Completions API. Its core idea is to package the complete sequence of reasoning, tool calls, web search, and final answer in an interaction into a traceable, resumable 'Response' object. Previously, when building agents with Chat Completions, developers had to implement the multiround tool-call loop themselves: send a request → receive a tool call → execute the tool → feed back the result → send another request. The Responses API brings that path together as 'give it a Response, then continue with the next Response,' automatically feeding back tool results and greatly simplifying client code. DeepSeek has not invented a new protocol here; it has adopted an existing format.

✓ Tool calls · built-in server-side web_search ✓ reasoning.effort (low/high/max) · structured text.format output ✓ Streaming SSE ends with response.completed / incomplete / failed ○ Stateless: previous_response_id / conversation / store not yet supported ○ file_search / code_interpreter / computer_use / mcp silently ignored

Put plainly, it does not replace Chat Completions, but opens a smoother path for agent and Codex scenarios in which 'AI does the work itself.' Existing Responses API clients can connect with almost no changes.

04 PRICE Pricing and openness: surprisingly cheap

Official prices (per million tokens, unchanged from Preview):

deepseek-v4-flash Input · cache hit: $0.0028 Input · cache miss: $0.14 Output: $0.28 per million tokens deepseek-v4-pro Input · cache hit: $0.003625 Input · cache miss: $0.435 Output: $0.87 per million tokens

For a 284B-parameter model, output costs about two yuan per million tokens; its cache-hit price ranks #1 of 101 on Artificial Analysis (about a 98% discount). DeepSeek also previewed time-of-day pricing: during peak hours (Beijing time 9:00–12:00 and 14:00–18:00), prices would double; the effective date was yet to be announced.

Its openness is similarly extensive: MIT license, 8-bit weights, 19 community quantized versions already available, and an official local-deployment configuration using four GB300 GPUs and vLLM. A Hacker News commenter said it was 'good enough to run locally on consumer hardware under $10,000.'

05 VOICE Community reaction: excitement with restraint

Hacker News: the day's hottest post (566 points / 275 comments). Some called it 'insane' that a roughly 200B-active-parameter model surpassed the 1.8T-parameter V4-Pro Preview on certain agent benchmarks. Many considered it good enough at a 50–100-fold cost difference. The main disappointment was the lack of multimodal/vision capabilities; others complained about confusing version names.

V2EX: positive but cautious. 'You can use it with Codex and GPT'; regarding the scores: 'The benchmarks look strong, but who knows how it performs in actual use.'

Zhihu: it climbed the trending list after release. A representative comment said, 'You really misjudged Liang,' expressing anticipation for the official V4-Pro release.

06 WRAP Conclusion: what is known and unknown

Known: official Flash public beta · much stronger agent capabilities · Responses API supported · MIT open source · unchanged prices Still to verify: official results on general benchmarks · effective date of time-of-day pricing · official V4-Pro release (expected in early August)

AI Writes About AI This article was written by DeepSeek-V4-Flash. Yes, by the very model whose official release arrived today. Its claims and figures are drawn from the official changelog, pricing page, Hugging Face model card, arXiv technical report, and public coverage, listed below.

Main sources api-docs.deepseek.com (changelog / pricing / Responses API guide) huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 arxiv.org/abs/2606.19348 · artificialanalysis.ai Discussions on Hacker News / V2EX / Zhihu

Rainy-Day Reflections · AI Observation · August 1, 2026