DeepSeek is retiring V4 Pro because their Flash model beat it

On 9 September DeepSeek told API users that their new V4.1 Flash model had "comprehensively surpassed V4 Pro across all key metrics, including performance, cost, speed, and task completion time", and that as a result every request to the Pro model would soon be routed to V4.1 Flash and billed at Flash prices. A day later they confirmed the V4.1 Flash launch and set the Pro cutover for 12:00 Beijing time on 14 September, which is 04:00 UTC.

That is an unusual thing for a model vendor to do. The normal pattern is to keep the expensive tier alive as long as anyone will pay for it. DeepSeek are doing the opposite: the moment their small model beat their large one, they stopped selling the large one. This post is about what that means if you run agents on either model, and what it does to the bill.

What is actually changing

Two things, on two dates.

• 10 September, 04:00 UTC. V4.1 Flash launched and the Flash price dropped. The deepseek-v4-flash model id now serves V4.1 Flash at the new rate.
• 14 September, 04:00 UTC. V4 Pro is discontinued. Requests to deepseek-v4-pro are routed to V4.1 Flash and billed at the Flash rate. Nothing changes in your code; the model id keeps working, you just get a different model behind it.

DeepSeek describe this as an interim arrangement "prior to the release of V4.1 Pro". So the Pro id is not going away, it is being held open for a model that does not exist yet. Until then, Pro is Flash.

The prices, on above.dev

All figures are per million tokens at cost plus ten percent, which is how every model here is priced. DeepSeek bill by time of day, so each row has an off-peak and a peak figure. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; everything else, including Beijing-time weekends, is off-peak. Those windows did not change.

Model	Input	Cache hit	Output
V4 Flash, until 10 Sep	$0.242 / $0.484	$0.0077 / $0.0154	$0.726 / $1.452
V4.1 Flash, from 10 Sep	$0.165 / $0.33	$0.0033 / $0.0066	$0.66 / $1.32
V4 Pro, until 14 Sep	$0.726 / $1.452	$0.0242 / $0.0484	$2.178 / $4.356
Pro id, from 14 Sep	$0.165 / $0.33	$0.0033 / $0.0066	$0.66 / $1.32

Flash got about a third cheaper. Pro got about four and a half times cheaper, because it stopped being Pro.

What it does to an agent session

Take the workload we see most on this gateway: a coding agent running for an hour, one million input tokens, 20,000 output tokens, with 90 percent of the input served from the prompt cache. Off-peak rates.

Model	Session cost
V4 Pro, last week	$0.138
V4.1 Flash, now	$0.033

Same model id if you were on Pro, same code, a quarter of the bill. The cache-hit rate is where most of the saving lives: it dropped from $0.0242 to $0.0033 per million, and in an agent loop cached tokens are most of the tokens.

The catch, and it is a real one

"Comprehensively surpassed" is DeepSeek's claim about their own model. We have not run our own evaluation of V4.1 Flash against V4 Pro, and you should not take a vendor's word for a claim that conveniently lets them retire their most expensive product. What we can say is narrower and still useful.

• If you are on Pro because you tested it and it works for your task, you have until 04:00 UTC on 14 September to test V4.1 Flash on the same task. After that the switch happens whether you tested or not.
• If your agent is tuned against Pro's specific behaviour, tool-call formatting, verbosity, how it handles long context, check those specifically. "Better on benchmarks" and "behaves the same in my loop" are different claims.
• If the new model is worse for you, there is no way to keep V4 Pro. DeepSeek are not offering a legacy endpoint. Your options on 14 September are V4.1 Flash, a different model, or waiting for V4.1 Pro with no date attached.

That last point is the part that deserves the word unusual. Most vendors deprecate with a long tail. DeepSeek gave five days from announcement to cutover, with a claim of superiority in place of a migration path.

What we did on the gateway

Both changes are scheduled in our pricing table rather than applied by hand, so billing flipped for Flash at exactly 04:00 UTC on the 10th and will flip for the Pro id at exactly 04:00 UTC on the 14th, with the old rates kept for anything billed before. The first announcement implied both would move on the 10th; the second corrected it. We had scheduled Pro for the 10th on the first reading and reversed it before any Pro traffic hit the window. Lesson for anyone else who resells DeepSeek: schedule only the leg the vendor has actually dated.

The catalogue label is the other thing to fix. From the 14th, a model listed as Pro at a higher capability tier will be serving Flash weights at Flash prices. We will collapse the two entries into one honest listing once the cutover has happened and we have seen it bill.

Try it before the 14th

Every response from above.dev carries an x-cost-usd header with the exact cost of that request. Run the same task against deepseek-v4-pro today and deepseek-v4-flash tomorrow, and you will have both the quality comparison and the cost comparison on your own workload before the decision is made for you.