The short version: DeepSeek announced on September 10 that its flagship V4 Pro model would stop answering its own requests from September 14, with everything routed to the cheaper V4.1 Flash. The coverage that followed wrote the obituary. As of today, September 15, DeepSeek’s live pricing page carries a footnote reversing it: “In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged.” Pro is still there, still billed at its own rates.
Two other models on that same page really were replaced, and nobody wrote about those. Footnote 1 says the names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them are retired and the requests are now served by V4.1 Flash. Your code keeps working. It gets a 200 back. A different model answers. The money involved in all of this, for a business-sized workload, is about 94 cents a month. The reason to care is not the money.
What actually changed on September 14?
The September 10 release notes for V4.1 Flash set out a transition with a specific clock on it. V4 Flash and V4 Flash Vision Exp were retired, their names kept alive as aliases pointing at the new model. And from 04:00 UTC on September 14, the notes said, all deepseek-v4-pro requests would route to V4.1 Flash at Flash rates, until a V4.1 Pro arrived.
That is the version that traveled, and it was a fair thing to write on September 10, because it was what the company had published. Then DeepSeek changed its mind. The footnote quoted above now sits on the pricing page and in the changelog entry for that same release, saying the opposite: Pro continues, billing unchanged, with further notice promised if that changes. The pricing table backs it up. Read today, deepseek-v4-pro still has its own column and its own numbers, at $0.66 per million input tokens off peak against Flash at $0.15, and $1.98 per million output against Flash at $0.60.
So the current state of play, verified on the vendor’s own pages on September 15: one flagship reprieved after users complained, two workhorse models quietly swapped out behind names that still resolve.
How much is the price difference actually worth?
Take a job an owner would recognize. A trades business drafts quotes from rough job notes, three hundred a month, feeding the model about 4,000 tokens of notes, price list and past-quote examples each time, and getting back about 800 tokens of draft. That is 1.2 million input tokens and 240,000 output tokens a month.
On Flash at off-peak rates, that costs 18 cents of input and roughly 14 cents of output. Call it 32 cents a month. On Pro, the same work costs about 79 cents of input and 48 cents of output, so roughly $1.27.
The gap is 94 cents. A 77 percent cut on input pricing and a 70 percent cut on output, applied to a real workload, comes to less than a dollar. That keeps happening: we found the peak and off-peak timing discount was worth 38 cents a month to an American business, and that two simultaneous price cuts from the largest providers moved a small business bill by cents.
Token prices are not where an owner’s AI money goes. It goes on the seat licenses of the software sitting on top of these models. Which is exactly why the substitution matters more than the saving.
Why does a silent model swap matter if the bill goes down?
Because the thing that changed is the part you tested.
If you or your bookkeeper spent a week last spring getting a prompt to produce quotes in your voice, with your line-item conventions and your terms at the bottom, that work was validation. You ran it, you read the output, you decided it was good enough to send to a customer. That decision was about one specific model.
V4.1 Flash is a different machine. It is a 552 billion parameter mixture-of-experts model that activates roughly 8 billion parameters for input and 16 billion for output, a fundamentally different amount of computation than a flagship applies. It may well be better. DeepSeek says testing showed it beating Pro on capability, cost and speed, and that is plausible. It is still not the thing you signed off on, and the swap happened without a code change, a version bump or an email.
For most owners this happens one level up, which makes it harder to see. You do not call DeepSeek. Your CRM does, or your quoting tool does, or the AI feature your booking software shipped in June. Your vendor made the model choice, and the model underneath that choice moved.
What should a small business owner actually do about this?
Not panic, and not learn to read API documentation. One question to your vendor covers it: which model does this feature use, and how will you tell me when it changes?
A good answer names a model and a version, and describes a notification. A vague answer, or a support rep who does not know, tells you something useful too. It tells you that the output quality you are relying on is not being monitored by anyone, which means the monitoring has to be yours.
The practical version takes ten minutes a month. Keep five or six real examples of good output from the tool, saved with the date, then run the same inputs again later and read what comes back. You are not benchmarking anything. You are checking that it still sounds like you, still gets your prices right, still puts the deposit terms in. This is the same discipline as checking the expiry date beside a price rather than the price itself, and it rests on the same fact: the AI layer of your business is being changed by other people on a schedule you do not control.
That is not a reason to stay away. The quoting workflow above costs about a third of a dollar a month in model fees and gives an owner back an evening a week. It is a reason to keep a hand on the wheel, because the judgement about whether the output is good enough to put your name on was always the part that belonged to you.
Frequently Asked Questions
Is DeepSeek V4 Pro still available?
Yes. DeepSeek announced on September 10 that V4 Pro requests would be routed to V4.1 Flash from September 14, but the company reversed that decision. A footnote on its pricing page and changelog, checked on September 15, 2026, states that API service for V4 Pro continues after that date with billing unchanged, in response to user demand. The pricing table still lists V4 Pro with its own separate rates.
What happens if my code still calls deepseek-v4-flash?
It keeps working, but a different model answers it. DeepSeek’s pricing page states that the names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted while the models behind them have been retired, with requests served by V4.1 Flash and billed at the Flash price. You get a normal successful response, so nothing in your logs will flag that the model changed.
How do I find out which AI model my software vendor is actually using?
Ask them directly, in writing, and ask two things rather than one: which model and version the feature runs on, and how they will notify you when that changes. Most small business software does not disclose this in its interface, so the support channel or your account manager is the route. If nobody at the vendor can answer, treat that as your answer and start keeping your own dated samples of the tool’s output.
Does the price difference between Pro and Flash matter for a small business?
Barely. Running three hundred AI-drafted quotes a month costs roughly 32 cents on Flash and roughly $1.27 on Pro at off-peak rates, a difference of about 94 cents. Token pricing at small business volume is close to a rounding error, and the real AI spend sits in the monthly per-seat fees of the software built on top of these models.
Has an AI feature in software you already pay for started behaving differently without any announcement? We would like to hear which tool it was.
