Summary: On May 22, 2026, DeepSeek announced that V4 Pro’s 75% discount had become permanent, converting the promotional quarter price into official pricing rather than extending a sale. The pricing page lists $0.87 per million output tokens, $0.435 input on cache miss, low cache-hit rates, and a 500-request concurrency cap. The move improves long-term budgeting and pressures rivals, while routing still has to account for capacity limits.
DeepSeek’s flagship, V4 Pro, is the subject of a May 22, 2026 announcement on X: its 75% discount “is now permanent.” The promotion had been scheduled to lapse at 15:59 UTC on May 31; instead of ending, official pricing moves to one quarter of the list price. Bloomberg picked the story up the following day under the headline “DeepSeek to Make Permanent 75% Discount on Flagship AI Model,” reading the move as pricing strategy rather than a one-off marketing push.
There is nothing routine about this cut. DeepSeek’s pricing documentation puts the discounted output rate at $0.87 per million tokens, and fixing it in place amounts to retiring the list price outright. For a lab built on open weights to promote a sale price into the official one is a message in itself.
A Sale With No Expiry Date
The pricing page says it without ceremony: the 75% V4 Pro promotion runs until May 31, 2026 at 15:59 UTC, and “after the promotion ends, Pro pricing will be officially adjusted to one quarter of the original price.” So this is not an extension of the sale — the discounted figure simply becomes the standing price. On the day the promotion lapses, the invoice looks exactly the same.
The developer response came fast. On Hacker News, a thread pointing directly at the pricing docs collected 621 points and 549 comments, and the conversation kept circling two questions: how long can a price this low last, and do competitors have to answer it?
What the Pricing Page Says
The May pricing page offers three sets of numbers worth holding onto:
- V4 Pro: $0.435 per million input tokens on cache miss, $0.87 output, cache hits as low as $0.003625, and a concurrency cap of 500
- V4 Flash: $0.14 input, $0.28 output, cache hits at $0.0028, and a concurrency cap of 2,500; both models ship with 1M context
- Since April 26, cache-hit pricing has sat at one tenth of the launch-price level
Run the discount backward and V4 Pro’s list output price comes to $3.48 per million tokens, of which the standing price keeps one quarter. Layer the one-tenth cache-hit rate on top and cache-heavy workloads see marginal cost shrink toward rounding error — for agents that pull the same system prompts and documents into context on every turn, that is a sweet spot engineered on purpose.
What Makes Permanence Sustainable
A discount turns permanent for one of three reasons: unit costs actually dropped, demand needs a jolt, or some mix of the two. All three readings fit DeepSeek. Open-weight labs lean on API revenue to build their ecosystem, which makes price the sharpest lever for winning developers — and an inference architecture with strong cache hit rates lets the unit economics hold even at these rates.
Supply structure tells the sharper story: Pro is capped at 500 concurrent requests against Flash’s 2,500. Flash — cheap and high-concurrency — absorbs the scaled traffic, while Pro holds the flagship label at a low price with supply deliberately reined in. That is capacity management, not charity. Teams running critical services on Pro will run into the concurrency cap long before the price ever becomes the constraint.
The Reading for Developers and Product Teams
For one, budgeting gets easier: the “will the promo end?” uncertainty disappears, and long-horizon cost models can be written straight against the standing price. For another, model routing gains strategic room — the output-price spread between Flash and Pro ($0.28 versus $0.87) is wide enough to carry a tiered design where routine work goes to Flash and hard problems escalate to Pro. And the move presses directly on commercial APIs in the same band: with an open-weights flagship selling output under $1 per million tokens, everyone else’s pricing cushion keeps thinning.
In our opening-of-year outlook we tracked how quickly model turnover and the open-source split were advancing through 2026; a discount made permanent is the latest data point plotted on that same curve. For product teams, the question has stopped being whether to use cheap models and become whether your routing architecture is built to take advantage of them.
The raw arithmetic deserves a line of its own: a workload producing one billion output tokens a month runs $870 at the standing Pro price versus $3,480 at list — a difference bigger than most optimizations an engineering team could land in a quarter. And because the price now carries official status instead of promotional status, the saving clears procurement review and finance sign-off rather than dying with a promo countdown.