DeepSeek released V4.1 Flash on September 10, 2026
DeepSeek’s newest open-weight model does not beat GPT-6 Astra at everything. It doesn’t have to. The bigger story is how close it can get on many everyday tasks while operating at a tiny fraction of the cost.
DeepSeek released V4.1 Flash on September 10, 2026, only days after OpenAI introduced GPT-6 Astra. On paper, Astra remains the more capable frontier model. But the economics surrounding the two models may be even more important than the benchmark crown.
DeepSeek V4.1 Flash is a 552-billion-parameter Mixture-of-Experts model with a novel Causal Encoder-Decoder architecture. Despite its enormous total parameter count, it activates only about 8 billion parameters while processing input and 16 billion during output generation. It supports a one-million-token context window, native vision and was pretrained on roughly 45 trillion tokens. DeepSeek has also released the weights under an MIT license.
That combination—frontier-level capabilities, sparse computation, extremely low API pricing and downloadable weights—is what should concern companies whose business models depend on expensive proprietary AI remaining significantly better than the alternatives.
The Cost Difference Is Enormous
OpenAI currently charges $10 per million input tokens and $50 per million output tokens for GPT-6 Astra. DeepSeek V4.1 Flash costs just $0.30 per million input tokens and $1.20 per million output tokens at peak pricing.
During DeepSeek’s off-peak hours, those numbers fall again to $0.15 and $0.60.
That means the raw API pricing looks roughly like this:
| DeepSeek V4.1 Flash | GPT-6 Astra | |
|---|---|---|
| Input | $0.30/M | $10/M |
| Output | $1.20/M | $50/M |
| Off-Peak Input | $0.15/M | $10/M |
| Off-Peak Output | $0.60/M | $50/M |
At peak prices, Astra is about 33× more expensive on input and 42× more expensive on output. Off-peak, the difference grows to roughly 67× and 83×.
So saying DeepSeek costs “around 50× less” is not an unreasonable shorthand, although actual application costs depend heavily on the ratio of input to output tokens, caching and how many reasoning tokens each model consumes.
There is another extraordinary difference: DeepSeek’s peak cached-input rate is just $0.006 per million tokens, compared with $1 per million cached tokens for Astra.
For large-scale agents constantly rereading context, repositories and documents, those numbers matter.
But Is DeepSeek Really as Good as Astra?
Not overall.
GPT-6 Astra remains considerably stronger when you look at the hardest frontier benchmarks. However Deepseek is very powerful.
Artificial Analysis currently gives V4.1 Flash an Intelligence Index score of 40, compared with 53 for GPT-6 Astra Max. Astra also has large advantages on Humanity’s Last Exam and Terminal-Bench 4.0.
But that isn’t the entire story.
On Artificial Analysis’ AutomationBench-AA evaluation, DeepSeek scored 69% versus Astra’s 68%. On GDPval-AA v2, DeepSeek scored 1,632 compared with Astra’s 1,580. Other evaluations similarly show that the performance gap changes substantially depending on what people are actually asking the models to do. Yes, Astra likely is more well rounded and capable… However, the cost difference is huge.
Web and design benchmarks show a gap, but perhaps not one large enough to justify the price difference for every customer.
On Arena’s WebDev leaderboard, GPT-6 Astra Max currently scores 1,800, while V4.1 Flash scores 1,614. In Brand & Marketing development tasks, Astra scores 1,747 while DeepSeek scores 1,628. Front-end development shows 1,812 versus 1,622.
Those are real Astra advantages. But consider the business question:
Would you pay 40 or 50 times more for a model that scores roughly 7–12% better on a task where either model produces a perfectly usable result? When
For extremely difficult research, cybersecurity, autonomous computer work or complex engineering, the answer may absolutely be yes.
For generating landing pages, analyzing marketing data, building ordinary applications, writing content, automating business processes and millions of other routine AI requests, the answer becomes much less obvious.
That is where DeepSeek becomes disruptive.
Was DeepSeek V4.1 Flash Also 50× Cheaper to Train?
This number requires more caution.
Neither company has published enough information to make a definitive dollar-for-dollar comparison of the total training cost of V4.1 Flash against GPT-6 Astra.
DeepSeek tells us that V4.1 Flash was trained from scratch on approximately 45 trillion tokens, and its sparse architecture activates dramatically fewer parameters than its headline 552-billion-parameter size would suggest.
OpenAI, meanwhile, has not publicly disclosed Astra’s full parameter count, total training compute or complete training bill in its launch materials.
My rough estimate is that the underlying cost difference could plausibly be measured in tens of times, and potentially approach 50× depending on what is included in the calculation. But until both companies disclose comparable compute and training-cost figures, 50× cheaper to train should be treated as an estimate—not an established fact. However, we know the training cost difference is substantial.
What is established is arguably just as important: DeepSeek has demonstrated that enormous capability can be produced with architectures that activate surprisingly little compute per token through innovation and hard work. And it is unlikely to stop here.
The Bigger Threat Is Open Source
Technically, V4.1 Flash is better described as an open-weight model rather than completely open-source AI because not every aspect of its training process and dataset is reproducible.
But commercially, the distinction may matter less.
The model weights are publicly available on Hugging Face under an MIT license, allowing companies and developers to run, modify and commercially deploy the model themselves.
That changes the competitive equation.
A proprietary AI company can lower its API prices. An open-weight model can eliminate the API provider entirely.
Companies can self-host it. Cloud providers can compete to serve it. Developers can optimize it for specialized hardware. Enterprises can fine-tune it around proprietary workloads. Countries can deploy it inside sovereign infrastructure without depending on an American AI provider.
And thousands of independent developers can improve the surrounding software simultaneously.
OpenAI and Anthropic are therefore competing against something larger than DeepSeek.
They are increasingly competing against the marginal cost of intelligence itself collapsing.
Why Regulation Has Suddenly Become Such an Important Issue
This makes the current AI-regulation debate particularly complicated.
OpenAI has recently called for mandatory, capability-based national AI safety requirements, while Anthropic has repeatedly advocated safety standards for increasingly powerful frontier systems.
There are legitimate reasons for those concerns.
OpenAI says GPT-6 Astra has reached its Critical cybersecurity capability threshold, meaning the model can potentially discover previously unknown vulnerabilities and develop sophisticated exploitation techniques when provided with appropriate tools and access. That represents a very different safety problem from the chatbots of only a few years ago.
It would therefore be unfair to simply claim that OpenAI or Anthropic wants regulation because regulation protects their businesses.
In fact, Anthropic CEO Dario Amodei recently stated explicitly that Anthropic does not support banning open-weight models and described open-weight systems without dangerous capabilities as a public good.
But there is nevertheless an unavoidable economic reality surrounding regulation.
The companies with the most to lose from unrestricted open-weight competition are also among the companies best equipped to comply with expensive regulation.
If advanced AI development requires enormous compliance departments, licensing procedures, safety testing regimes and regulatory approval, billion-dollar frontier laboratories can absorb those costs far more easily than startups and open-source projects.
That does not make regulation wrong.
It does mean regulators must be extremely careful that AI safety does not accidentally become AI regulatory capture.
OpenAI and Anthropic Have a Real Business Problem
The immediate threat isn’t that DeepSeek V4.1 Flash suddenly makes GPT-6 Astra obsolete.
It doesn’t.
Astra is still substantially better at several of the hardest things AI systems can currently do.
The threat is the trajectory.
Imagine the performance gap continuing to shrink while the cost difference remains 20×, 40× or 50×.
Then imagine another open-weight model arriving six months later that closes half the remaining gap.
Then another.
At some point, the question changes from:
“Which company has the smartest model?”
to:
“How much smarter does a proprietary model need to be before customers will pay 50 times more for it?”
That is a much more dangerous question for OpenAI and Anthropic.
The Real AI Race May Be Cost Per Unit of Intelligence
For the last several years, the AI race has been described primarily in terms of intelligence: who has the highest benchmark score, the biggest model or the most impressive demonstration.
DeepSeek V4.1 Flash illustrates why the next stage may be different.
The important metric could increasingly become:
How much useful intelligence can you deliver for one dollar?
By that measurement, open-weight AI is advancing extraordinarily quickly. You can only monopolize a market so much when similar options are available at a much cheaper price.
GPT-6 Astra may be the superior model today. But V4.1 Flash doesn’t need to defeat Astra outright to disrupt OpenAI’s economics.
It only needs to become good enough for 90% of what customers actually do while being dramatically cheaper for them to operate.
And if the rate of improvement in open-weight models continues at anything close to its current pace, the economic moat protecting the world’s largest closed AI laboratories could shrink much faster than most people expect.