DeepSeek V4 Flash 0731 officially launched on July 31, 2026. It costs $0.14 per million input tokens. It scores 82.7 on Terminal-Bench 2.1. It beats its own Pro model on agentic coding tasks. And it costs 7x less than GPT-5.6 Luna and 21x less than Claude Sonnet 5. This is the full review of whether the hype is real, who should use it, and what the catch is.
DeepSeek V4 Flash is a sparse Mixture-of-Experts AI model from DeepSeek. It has 284 billion total parameters but only activates 13 billion per token during inference, making it fast and cheap to run. The 0731 version launched July 31, 2026 with re-post-training that dramatically improved its coding and agent benchmarks. It costs $0.14 per million input tokens, supports a 1 million token context window, and is available via the DeepSeek API and OpenRouter. For most developers who need a fast, affordable coding assistant or agent backbone, it is currently the best value model available in 2026.
- What is DeepSeek V4 Flash?
- What changed in the 0731 update?
- Benchmark results โ the real numbers
- Pricing โ how cheap is $0.14/M really?
- Key features
- DeepSeek V4 Flash vs competitors
- Pros and cons
- Who should use DeepSeek V4 Flash?
- How to use DeepSeek V4 Flash
- Privacy and data concerns
- Frequently asked questions
- Final verdict and rating
What Is DeepSeek V4 Flash?
DeepSeek V4 Flash is an AI language model built by DeepSeek, a Chinese AI research lab backed by the hedge fund High-Flyer. It uses a Mixture-of-Experts architecture, which means the model has 284 billion parameters in total but only activates 13 billion of them for any given input. This makes it dramatically cheaper to run than a dense model of similar quality, which is why the price can be as low as $0.14 per million tokens.
The "Flash" in the name signals its positioning: fast inference, low cost, high throughput. It is not DeepSeek's most capable model โ that is V4 Pro โ but after the 0731 update, it now outperforms V4 Pro on the coding benchmarks that most developers actually care about. That is a genuinely unusual situation in the AI industry and is the main reason this release is getting so much attention.
The model is MIT-licensed, which means the weights are public and can be self-hosted. You are not locked into DeepSeek's API. If your use case involves sensitive data, you can run the model entirely on your own infrastructure with zero data leaving your environment.
๐ง In plain terms: Think of DeepSeek V4 Flash as a very smart specialist. It does not handle every task equally well, but for coding, agentic workflows, and reasoning tasks, it punches well above its price point. At $0.14 per million input tokens, it costs about the same as a cup of tea to process one million words of input.
What Changed in the 0731 Update?
The April 2026 preview of V4 Flash was capable but unimpressive on agent benchmarks. The 0731 version is a different story. DeepSeek applied what they call "re-post-training" โ which means they kept the same base model architecture and weights, but ran a new post-training cycle with better data and techniques specifically targeting coding and agent tasks.
The results are striking. Terminal-Bench 2.1 jumped from 56.9 to 82.7 โ a 25.8-point improvement. DeepSWE, which tests real-world software engineering ability, went from 7.3 to 54.4 โ a 7x gain in a single post-training cycle. These are not marginal improvements. They represent a fundamentally different level of usefulness for software engineering workflows.
Three new capabilities also shipped with the 0731 release:
- Native Responses API support. The model now speaks the Responses API protocol natively, making it a drop-in replacement for OpenAI Responses API users.
- Codex compatibility. V4 Flash now works natively with OpenAI Codex, opening it up to the large ecosystem of developer tools built on the Codex protocol.
- DSpark speculative decoding. The DSpark module is baked into the checkpoint, which improves inference speed without requiring separate configuration.
โฐ Timing note: DeepSeek released V4 Flash 0731 one day after OpenAI cut GPT-5.6 Luna's price by 80%, bringing it to $0.20 input and $1.20 output. DeepSeek's pricing at $0.14 and $0.28 was already lower and remained unchanged. The 0731 release was about capability improvements, not a price war response โ though the timing made it look that way.
Benchmark Results โ The Real Numbers
These are the benchmarks that matter for developers evaluating DeepSeek V4 Flash for real workloads. We have noted where scores are vendor-reported versus independently verified.
โ ๏ธ Benchmark caveat: Terminal-Bench 2.1 and DeepSWE scores above are vendor-reported by DeepSeek. Independent labs have not yet published fully replicated results for V4 Flash 0731. Treat these numbers as directional signals rather than final verdicts. The Artificial Analysis Intelligence Index score of 50 is independently verified.
Pricing โ How Cheap Is $0.14/M Really?
The headline price is $0.14 per million input tokens. To understand what that actually means, here are some real-world calculations.
- Processing 1,000 average developer messages (around 500 tokens each) costs roughly $0.07 โ seven cents.
- Running 10,000 coding agent calls per day at 800 tokens average input costs about $1.12 per day or $33 per month.
- The cached input price of $0.0028 per million tokens means repeated-context workloads โ like systems prompts or document contexts โ cost nearly nothing after the first call.
๐ก Peak pricing warning: DeepSeek has announced plans to introduce peak-hour pricing at 2x normal rates for 7 hours per day. The schedule has not been officially confirmed yet. If you are planning production workloads, account for the possibility of $0.28/M input and $0.56/M output during peak hours.
Key Features of DeepSeek V4 Flash 0731
Mixture-of-Experts Architecture
The MoE design is why V4 Flash is so affordable. Instead of running all 284 billion parameters on every token, the model uses a routing mechanism to activate only the 13 billion parameters most relevant to each input. This reduces compute cost dramatically without sacrificing quality on the tasks the active experts specialize in. For coding and reasoning, those experts are highly trained, which is why the benchmark scores are competitive with much more expensive models.
1 Million Token Context Window
The 1,048,576 token context window is one of the largest available in any model at this price point. In practice, that is enough to hold an entire medium-sized codebase, a full legal contract corpus, or hours of transcript in a single inference call. The maximum output of 384,000 tokens is also unusually high, making V4 Flash well-suited for tasks that require generating large amounts of structured content.
Codex Compatibility and Responses API
The 0731 release added native support for both the OpenAI Responses API and Codex protocol. This is practically significant. It means developers using Codex-compatible tools like Cursor, Claude Code integrations, and dozens of other coding environments can now route requests to V4 Flash with a simple model name change โ no other code modifications needed.
MIT License and Self-Hosting
The model weights are publicly available on Hugging Face under an MIT license. You can download them and run the model entirely on your own hardware. For enterprises with data residency requirements, healthcare organizations, or any team that cannot send data to Chinese servers, self-hosting eliminates the privacy concern entirely while keeping the cost advantage of the MoE architecture.
DeepSeek V4 Flash vs Competitors
| Model | Input Price | Output Price | Context | Terminal-Bench | License | Best For |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash 0731 | $0.14/M | $0.28/M | 1M | 82.7% | MIT | Coding, agents, high-volume |
| GPT-5.6 Luna | $0.20/M | $1.20/M | 1.05M | Lower | Proprietary | Fast, general tasks |
| GPT-5.6 Terra | $2.50/M | $15/M | 1.05M | Higher | Proprietary | Balanced production |
| Claude Sonnet 5 | ~$3/M | ~$15/M | 1M | 63.2% | Proprietary | Daily use, writing, coding |
| Claude Opus 4.8 | ~$15/M | ~$75/M | 200K | ~85% | Proprietary | Complex reasoning |
| DeepSeek V4 Pro | $0.43/M | $1.20/M | 1M | 72.1% | Proprietary | Higher reasoning tasks |
The key takeaway from this table: DeepSeek V4 Flash 0731 is the only model in this list that beats models costing 3x to 21x more on Terminal-Bench 2.1. It is not the best at everything, but for agentic coding workflows it currently offers the best benchmark-per-dollar of any publicly available model.
Pros and Cons
Who Should Use DeepSeek V4 Flash?
How to Use DeepSeek V4 Flash in 2026
Privacy and Data Concerns โ What You Need to Know
This is the part of most DeepSeek reviews that gets glossed over. It should not be.
When you use DeepSeek's hosted API, your prompts and outputs are processed on servers located in China. DeepSeek is a Chinese company subject to Chinese law, including requirements to cooperate with government data requests. This is not speculation โ it is a factual description of the legal environment any Chinese technology company operates in.
For many use cases, this is not a problem. If you are building a public-facing chatbot, processing non-sensitive content, or doing rapid prototyping, the risk profile is low. But there are categories of use where this matters significantly:
- Proprietary source code or business logic
- Healthcare data subject to HIPAA or similar regulations
- Financial information covered by data residency rules
- Personal data of EU citizens covered by GDPR
- Government or defense-adjacent work
The solution for sensitive workloads is self-hosting. The MIT license makes this fully legal and straightforward. A team with 13B active parameters worth of GPU capacity can run V4 Flash entirely within their own environment. The model's MoE architecture actually makes self-hosting more accessible than a similarly performing dense model would be, since only 13 billion parameters are active at inference time.
๐ Our recommendation: For public, non-sensitive workloads, use the hosted API freely. For any data that has legal, regulatory, or commercial sensitivity, self-host the MIT-licensed weights. Do not send regulated data to any hosted API you do not control โ this applies to GPT, Claude, and Gemini as well as DeepSeek.
Our Rating
Frequently Asked Questions
DeepSeek V4 Flash is a sparse Mixture-of-Experts AI model with 284 billion total parameters and 13 billion active parameters per inference. The 0731 version launched July 31, 2026 after re-post-training that dramatically improved coding and agent benchmarks. It costs $0.14 per million input tokens, supports a 1 million token context window, and is MIT-licensed for self-hosting. It is best suited for coding, reasoning, and agentic workflows where cost efficiency is a priority.
DeepSeek V4 Flash 0731 costs $0.14 per million input tokens on a cache miss, $0.0028 per million tokens on a cache hit (a 98 percent discount), and $0.28 per million output tokens. The blended effective rate for a workload with a 70 percent cache hit rate is approximately $0.06 per million tokens. DeepSeek plans to introduce peak pricing at 2x normal rates for 7 hours per day, though the final schedule has not been announced.
On Terminal-Bench 2.1, DeepSeek V4 Flash 0731 scores 82.7 versus GPT-5.6 Luna's lower score on the same benchmark. It costs $0.14 per million input tokens versus Luna's $0.20 and $0.28 output versus Luna's $1.20, making Flash significantly cheaper on output. For agentic coding tasks specifically, Flash currently beats Luna on benchmarks at lower cost. For general intelligence, writing, and tasks outside coding, GPT-5.6 Luna has a stronger general capability profile.
The architecture and total parameter count stayed the same. DeepSeek applied re-post-training with improved data targeting coding and agent tasks. Terminal-Bench 2.1 jumped from 56.9 to 82.7, a 25.8-point increase. DeepSWE improved from 7.3 to 54.4, a 7x gain. New features added include native Responses API support, Codex compatibility, and the DSpark speculative decoding module baked into the checkpoint.
The hosted DeepSeek API sends data to servers in China. DeepSeek is a Chinese company subject to Chinese law. For non-sensitive, public-facing workloads this is generally acceptable. For regulated data including healthcare, financial, personal data of EU citizens, or proprietary code, use the self-hosted version instead. The model weights are MIT-licensed and freely downloadable from Hugging Face, so self-hosting is a straightforward option for privacy-sensitive use cases.
There is no free tier for the hosted API. However, the pricing is so low that it functions similarly. A developer making 1,000 API calls per day at 500 tokens average input spends roughly $0.07 per day. For truly zero-cost access, download the MIT-licensed weights from Hugging Face and self-host โ you pay only your infrastructure costs, not per-token fees.
Use the DeepSeek API with the model ID deepseek-v4-flash. The API is OpenAI-compatible: change your base URL to DeepSeek's endpoint and update the model name. No other code changes are required. Alternatively, access it through OpenRouter with 22 providers for better uptime, or download the weights from Hugging Face for self-hosting. The model natively supports the Responses API and Codex protocol as of the 0731 release.
V4 Flash is optimized for speed, cost, and high-throughput workloads. V4 Pro is DeepSeek's higher-capability flagship. The 0731 update to Flash produced a surprising result: Flash now outperforms Pro on Terminal-Bench 2.1 at 82.7 versus Pro's 72.1. V4 Pro still leads on the overall intelligence index and complex reasoning tasks. V4 Flash is cheaper at $0.14/M input versus V4 Pro at $0.43/M input. For most coding and agent tasks, Flash is now the better choice.
No. DeepSeek V4 Flash 0731 is a text-only model. It accepts text input and produces text output. It does not support image, audio, or video input in this release. For multimodal tasks, consider GPT-5.6 Sol or Claude Sonnet 5, which both support image input.
It depends on your use case. If your primary workload is coding, automation, or agentic tasks and cost is a concern, switching makes sense: V4 Flash scores 82.7 on Terminal-Bench versus Sonnet 5's 63.2, at 21x lower cost. If your workload includes writing, research, creative tasks, customer-facing conversations, or multimodal inputs, Claude Sonnet 5 delivers a more consistent, polished experience. Many teams use both: V4 Flash for high-volume coding pipelines, Sonnet 5 for quality-sensitive or general-purpose tasks.
Final Verdict
DeepSeek V4 Flash 0731 is the most cost-efficient frontier-class AI model available in August 2026. That statement is not close. At $0.14 per million input tokens with an 82.7 Terminal-Bench score, it beats models costing 3x to 21x more on the benchmark that matters most to developers building coding agents and automation workflows.
The privacy concern is real and should not be dismissed. If your data is sensitive, regulated, or proprietary, use the MIT-licensed self-hosted version. If your data is non-sensitive, the hosted API is a legitimate choice that will dramatically reduce your AI infrastructure costs.
The 0731 benchmark improvements are vendor-reported and not yet independently verified. Treat the specific numbers as directional. The direction is clear: this model is significantly more capable for coding tasks than the April preview, and the independent Intelligence Index score of 50 confirms it performs well above average for its size class.
For most developers, our recommendation is straightforward. Add V4 Flash to your stack for coding agents and high-volume pipelines. Keep Claude Sonnet 5 or GPT-5.6 for general reasoning, writing, and customer-facing conversations. Running two models based on task type is not a compromise โ it is the smart engineering decision in 2026.
๐ Explore more on Find Best AI: Browse 500+ AI Tools โ | Compare AI Models โ | Claude Fable 5 Guide โ | GPT-5.6 Sol Review โ