DeepSeek launches V4‑Flash API with post‑training boost and low‑cost pricing
DeepSeek released the public beta of its DeepSeek‑V4‑Flash API on 31 July 2026. The model keeps the same architecture and parameter scale as the earlier V4‑Flash preview, but has been re‑post‑trained, which the company says drives markedly higher agent‑oriented benchmark scores.
In internal tests the V4‑Flash‑0731 achieved 82.7 on Terminal Bench 2.1, 76.7 on Cybergym, and a 1586 score on the Frontend Code Arena (ranked #7 overall, #3 among open‑weight models). The model also supports the Responses API natively and is optimized for OpenAI Codex, easing integration into AI‑agent applications.
Pricing is set at $0.14 per million input tokens and $0.28 per million output tokens, roughly one‑third the cost of DeepSeek’s V4‑Pro offering. DeepSeek announced that a V4‑Pro API will follow “as soon as possible.” The release adds a cost‑effective, high‑performance option to the competitive generative‑AI market.
Entities
DeepSeek · DeepSeek Co., Ltd. · DeepSeek-V4-Flash · DeepSeek-V4-Pro · DeepSeek‑V4‑Flash · DeepSeek‑V4‑Pro · July 31 2026 · OpenAI Codex · Responses API · Terminal Bench 2.1
Claims
What the coverage asserts, and how well corroborated each claim is across sources.
- [● 3 SOURCES] Pricing for the V4‑Flash API is $0.14 per million input tokens and $0.28 per million output tokens. donanimgunlugu.com · api-docs.deepseek.com · tech-noisy.com
- [● 6 SOURCES] DeepSeek‑V4‑Flash‑0731 retains the same architecture and parameter scale as the preview version. www.b2bnn.com · www.igeekphone.com · www.leiphone.com · techgenyz.com · api-docs.deepseek.com · +1 more
- [● 9 SOURCES] DeepSeek released the official DeepSeek‑V4‑Flash API public beta on July 31, 2026. www.b2bnn.com · www.igeekphone.com · www.leiphone.com · techgenyz.com · api-docs.deepseek.com · +3 more
- [● 5 SOURCES] DeepSeek‑V4‑Flash includes native support for the Responses API and is optimized for OpenAI Codex. www.igeekphone.com · techgenyz.com · donanimgunlugu.com · api-docs.deepseek.com · www.leiphone.com
- [● 3 SOURCES] DeepSeek performed post‑training (re‑post‑training) on the V4‑Flash model without changing its architecture, leading to higher agent benchmark scores. www.b2bnn.com · www.leiphone.com · techgenyz.com
- [● 4 SOURCES] In the Terminal Bench 2.1 benchmark DeepSeek‑V4‑Flash scored 82.7, surpassing the V4‑Pro preview. donanimgunlugu.com · www.igeekphone.com · www.leiphone.com · techgenyz.com
- [● 4 SOURCES] In the Cybergym benchmark DeepSeek‑V4‑Flash scored 76.7, surpassing the V4‑Pro preview. donanimgunlugu.com · www.igeekphone.com · www.leiphone.com · techgenyz.com
- [● 2 SOURCES] DeepSeek announced that a V4‑Pro API will be released “as soon as possible.” www.igeekphone.com · api-docs.deepseek.com
- [● 5 SOURCES] DeepSeek‑V4‑Flash outperformed the DeepSeek‑V4‑Pro preview across multiple agent benchmarks. thetechnologyexpress.com · www.b2bnn.com · www.igeekphone.com · cryptobriefing.com
- [● 4 SOURCES] DeepSeek‑V4‑Flash scored 82.7 on Terminal Bench 2.1 and 76.7 on Cybergym. www.b2bnn.com · cryptobriefing.com · www.leiphone.com
- [● 2 SOURCES] DeepSeek‑V4‑Flash has 284 billion total parameters with only 13 billion active during inference. donanimgunlugu.com · tech-noisy.com
- [● 5 SOURCES] DeepSeek‑V4‑Flash‑0731 retains the same architecture and parameter scale as the earlier V4‑Flash preview. www.b2bnn.com · api-docs.deepseek.com · www.igeekphone.com · tech-noisy.com · www.leiphone.com