DeepSeek has returned to the centre of the global AI price war with DeepSeek-V4-Flash-0731, an upgraded open-weight model released on July 31, 2026. The Chinese startup is not claiming the highest score across every test. Instead, it is challenging rivals with a sharper proposition: strong coding, reasoning, and agent performance at a remarkably low operating price.
The timing is important because businesses are moving from small pilots to products that generate thousands or millions of daily responses. Research firm Artificial Analysis estimates that the model costs about three cents to complete its Intelligence Index evaluation workload. That figure has quickly drawn attention from developers, startups, and larger companies searching for ways to run AI tools at scale without letting inference bills climb uncontrollably.
What DeepSeek Launched And Why It Is Trending
DeepSeek-V4-Flash-0731 is the official release of the V4-Flash model that first appeared as a preview in April. DeepSeek says the July update keeps the same architecture and model size but adds fresh post-training aimed at much stronger agent performance. The company’s official X announcement said the API had entered public beta, while a follow-up post clarified that the upgrade did not increase the model’s size.
Its specifications help explain the efficiency claim. The model contains 284 billion parameters, but only 13 billion are active during inference. It also supports a one-million-token context window, allowing developers to process long documents, large codebases and extended conversations. DeepSeek has released the weights under an MIT licence through its official Hugging Face model page, giving businesses broad commercial-use rights.
Key details behind the launch include:
- API pricing of $0.14 per million input tokens and $0.28 per million output tokens.
- An estimated evaluation cost of $0.03 on Artificial Analysis’ benchmark workload.
- An Intelligence Index score of 50, matching several faster, higher-priced mainstream models.
- Improved results across coding agents, repository work, cybersecurity tasks and tool-use benchmarks.
- Open model weights, commercial use through the MIT licence and support for a one-million-token context.
How Cheap Is DeepSeek V4 Flash Compared With Rivals?
The headline comparison comes from Artificial Analysis, which calculates the cost required for models to complete its Intelligence Index workload. DeepSeek-V4-Flash-0731 averaged roughly $0.03 per test. Moonshot AI’s Kimi K3 cost about $0.86, OpenAI’s GPT-5.6 Sol cost $1.86, and Anthropic’s Claude Fable 5 cost $3.15 on the same evaluation method. That makes the Anthropic model more than 100 times costlier for this particular test.
Price does not mean DeepSeek has defeated every premium model. Its score of 50 places it below some frontier systems, and Artificial Analysis reported an 84% hallucination rate on its AA-Omniscience testing despite an improvement from the earlier version. Buyers therefore need to compare accuracy, reliability, speed, data policy and task quality rather than choosing a model only because its token price looks attractive.
Still, the gap is large enough to alter buying decisions. A customer-support platform processing millions of routine requests could route basic summaries, extraction work and code checks to a lower-priced model, then reserve expensive systems for complex decisions. That multi-model approach is becoming one of 2026’s biggest enterprise AI trends as companies seek lower costs without relying on one provider.
Why This Release Could Push The AI Price War Further
DeepSeek’s launch arrives during a crowded week for Chinese AI. Alibaba introduced Qwen3.8-Max, while Moonshot, Z.ai, MiniMax and ByteDance continue releasing capable models aimed at developers worldwide. The competition is no longer limited to benchmark rankings. Providers are now fighting over API pricing, open weights, context length, coding agents and deployment flexibility.
For startups, cheaper inference can stretch limited funding. For established software companies, it can improve margins on AI features that users expect but may not pay extra to access. Open weights also let technical teams host the model themselves, customise deployments and reduce dependence on a single cloud vendor.
There are trade-offs. Self-hosting a 284-billion-parameter model still requires expensive hardware, skilled engineers, security controls and ongoing monitoring. Political filtering, data governance and regional compliance may also affect adoption. DeepSeek’s low API rates remove one major cost barrier, but they do not remove the work required to operate AI safely.
The broader message is clear: capable AI is becoming cheaper faster than many businesses expected. DeepSeek-V4-Flash-0731 may not be the smartest model available, yet its pricing could pressure OpenAI, Anthropic, Google and Chinese rivals to offer lower-cost tiers or more efficient models.
DeepSeek V4 Flash FAQs
1. What Is DeepSeek-V4-Flash-0731?
It is DeepSeek’s open-weight AI model designed for reasoning, coding, agents and long-context tasks worldwide.
2. How Much Does DeepSeek V4 Flash Cost?
DeepSeek charges $0.14 per million input tokens and $0.28 per million generated output tokens currently.
3. Is DeepSeek V4 Flash The Best AI Model?
No, several premium models score higher, but DeepSeek offers strong performance for its operating price.
4. Can Businesses Use The Model Commercially?
Yes, its MIT licence permits commercial use, modification and self-hosting under the licence terms freely.
5. Why Are Researchers Calling It The Cheapest Major Model?
Artificial Analysis measured roughly three cents per benchmark test, far below costs recorded for rivals.
