A design task that runs $1.61 on GPT-6 Astra costs about 2.3 cents on DeepSeek’s new model. Same brief, same output, one bill sixty times smaller. That single number is the whole pitch.
DeepSeek widened that gap on purpose, and it did it three days after teaching its chatbot to talk out loud.
Vision is the other half of that release cycle, and it’s worth a look on its own: whether the cheap new vision model is actually any good.
How much cheaper are we talking?
On September 10, DeepSeek released V4.1-Flash. It is a 552-billion-parameter model, but it only switches on a small slice of itself for each request. Eight billion parameters wake up to read your prompt. Sixteen billion handle the reply. And that is where the savings come from.
The public price list is blunt. Off-peak, DeepSeek charges 15 cents per million input tokens and 60 cents per million output tokens. GPT-6 Astra charges $10 and $50 for the same units. Run that mix and you land around 60 times cheaper.
Cache hits push it further. Send the same context twice, and DeepSeek bills about 0.3 cents per million tokens, which turns repeated prompts into something close to free. That is not a discount. That is a rounding mistake you can build a business on.
A third party, Artificial Analysis, ran a more realistic test. It put the cost of a single agent task at about 27 cents on DeepSeek, less than a thirteenth of what GPT-6 or Claude Fable 5.1 charged for the same job. On a design contest run by OpenDesign, DeepSeek finished for 2.3 cents against $1.61 for GPT-6.
The off-peak idea deserves its own beat. DeepSeek charges full price during busy weekday windows and half price when the servers are quiet. It is a nudge, not a rule. Batch your work into the middle of the night and the same tokens cost you less.
There is a second lever. DeepSeek publishes the weights under an MIT license, so a company can download the model and run it on its own hardware. That removes the per-token bill entirely, at the cost of buying GPUs. Most readers will never do this, and the point is that it is allowed.
DeepSeek has not always moved this way. In August it announced a price increase, and I wrote about why that call made sense at the time. Three weeks later, it is cutting instead.
And it talks now?
Two days after the price cut, DeepSeek did something quieter. On September 12, its phone app began testing voice chat with a small group of users. A tiny speaker icon shows up in the corner. Tap it, talk, and the app answers with a voice of its own.
The voices have names. Shell is playful. White Wave sounds steady. Starfish is sweet.
And Dark Tide is low and deep. Two men, two women, picked for different moods instead of padding out a feature list.
But the voice part barely made a ripple in English tech news. It ran in Chinese outlets first. If you live outside China, the odds are good you never saw a word about it.
DeepSeek has handled voice input since January, but only in one direction. You spoke, it transcribed, and it answered on screen. The new build closes the loop.
Its models already run inside Tesla cars, Huawei watches, and Xiaomi speakers, where the hardware borrowed DeepSeek’s brain but used its own voice. This is the first time DeepSeek speaks for itself.
Voice is the part Chinese users have been waiting on. At home, DeepSeek counts about 130 million monthly users. That trails Doubao and Alibaba’s Qwen. A talking assistant is table stakes in that market, and DeepSeek was late to it.
But the test is small. That matters. A grayed-out rollout means the company is still tuning the voices and watching for crashes. Four voices two days after a model launch is a signal, not a finished product.
Most people will never touch an API bill, and that is fine. Cheap tokens still reach you. They show up as cheaper pricing inside apps that quietly route their bulk work to whichever model costs least. Features that once felt too expensive to give away get easier to ship.
None of this makes V4.1-Flash the better model overall. It loses to GPT-6 on the hardest agent benchmarks, and its score on Terminal-Bench 3.0 sits near 30. Long documents trip it up too.
One tester fed it a 100,000-character novel and watched the same event happen a different number of times in different chapters. Another checked its stock report. The summary said 15.1% annual return. The chart said 17.8%.
I keep going back and forth on whether I trust it for real work. Cheap tokens are lovely until a wrong number slips into something I sent a client.
I read the whole announcement on my phone, in my kitchen, with coffee going cold on the counter. That is probably why the 2.3-cent figure stuck with me.
GPT-6 Astra is not standing still, either. OpenAI shipped it on September 3 and calls it the most capable model it has broadly released. It answers hard questions faster and holds a million tokens of context. You pay for all of that.
The catch is different in the US. The voice feature lives inside a Chinese app, behind a slow rollout most people will never see. Off-peak pricing is easiest to use if you schedule your jobs for the middle of the night. That is work you have to own.
The honest answer is that you route, you do not switch. Put the boring, high-volume steps on the cheap model. Keep the hard reasoning on the expensive one. The teams getting the most out of this month already ask which model fits each step, not which model wins.
For a company that built its name on quiet releases, a talking assistant is a strange fit. It is also a clear sign that DeepSeek wants to be an app you open every day, not just an API developers call at night.
If you do try the API, expect a few rough edges. A stack of odd error codes showed up in my first week, and I wrote up every one of them.
So the question is no longer whether a cheap model can keep up. On the tasks most people actually run, it can. The question is what you do with the budget you just freed up.
