DeepSeek's V4-Pro went GA in silence, then got a lot more expensive

DeepSeek's biggest model update of the summer arrived with no press release, no blog post and no changelog entry. On 12 August 2026 the company changed a single line on its own API pricing page: the deepseek-v4-pro endpoint would now route to a new checkpoint, DeepSeek-V4-Pro-0813, closing out a preview that had run since 24 April. Four days later the consequential news landed — a pricing restructure that raised the model's rates by as much as 1,100%, depending on token type and time of day. Read together, the silence and the surge tell a more useful story about DeepSeek's position than either fact does alone.

DeepSeek has a well-documented habit of shipping without ceremony, but the 0813 rollout took that further than usual. There was no dedicated announcement page and no changelog entry — the change surfaced only because outlets tracking the pricing page noticed the version string had moved.

That ambiguity produced genuinely different GA dates in the coverage. Outlets tracking the pricing-page swap itself treat 12 August as the operative date. Enterprise DNA and others date general availability to 13 August, when the change was first widely reported and confirmed as a production checkpoint rather than a continuing preview.

The distinction matters less for the calendar than for what it reveals about intent. A lab confident in a launch typically wants credit for it — a keynote, a benchmark chart, a pricing announcement timed for attention. DeepSeek instead let a four-month-old preview quietly graduate, and only volunteered the real news — the price increase — once it could no longer be avoided, four days later.

At launch, V4-Pro billed at a flat $0.435 per million input tokens, $0.87 per million output tokens, and $0.003625 per million tokens on a cache hit — pricing DeepSeek had made permanent in May after an earlier promotional cut. That flat rate did not survive general availability.

DeepSeek frames off-peak as a 50% discount, and mechanically it is — off-peak is exactly half the peak rate. But the discount is measured against the new peak price, not the old flat one. Even the cheaper off-peak tier — $0.66 input, $1.98 output — sits well above the $0.435/$0.87 that held from May to mid-August. There is no hour of the day where the price returns to what it was in July.

DeepSeek's own model card reports large jumps over the April preview: DeepSWE up from 12.8 to 62.7, CyberGym from 52.7 to 83.3, Terminal Bench 2.1 from 72.1 to 87.9. On the number that headline is built from, DeepSWE, that is a gain of nearly 50 points in four months.

None of those three figures had independent replication as of mid-September. Worse, where a neutral third party did check, the gap was not small.

MindStudio's independent read of the vendor figures still ranks V4-Pro competitively against named rivals on that specific Terminal-Bench 2.1 number.

The two data points are not contradictory so much as measuring different things — one compares DeepSeek's own number against rivals' own numbers, the other checks DeepSeek's number against a neutral harness. On the neutral harness, V4-Pro drops well below several rivals it beats on the vendor chart. Every lab loses points moving to a stricter reference harness; DeepSeek's fall is large enough to be an outlier rather than noise.

Set the vendor card aside and there is a genuinely strong, independently checked result. Vals.ai ran V4-Pro-0813 through its own SWE-bench Verified pipeline — a test of fixing real GitHub issues — and placed it second among 82 models at 96.40%, behind only Claude Opus 5's 97.00%, ahead of GPT-5.6 Sol and Grok 4.6, at around $0.02 per test against Opus 5's $1.29. DeepSeek never claimed a SWE-bench number itself, which is precisely why this result carries more weight than the vendor card.

Even at the new peak rate, $3.96 per million output tokens is nowhere near frontier closed pricing — reporting elsewhere puts OpenAI's GPT-5.6 Sol at $30 per million output, an order of magnitude higher. The gap that made DeepSeek's models attractive to cost-conscious builders has narrowed sharply — output pricing alone moved 2.3x off-peak and 4.5x at peak — but the ranking has not changed.

The pricing has already moved once more since the restructure. DeepSeek initially signalled it would retire the deepseek-v4-pro endpoint in favour of a newer V4.1 Flash architecture on 14 September, then reversed course after demand for the older model held up.

DeepSeek said on its API pricing page on 11 September that it had decided to carry on providing API services for DeepSeek V4 Pro beyond 14 September 2026.

No -0813-tagged weights have been published on Hugging Face, so the checkpoint behind the benchmark numbers cannot yet be independently reproduced by self-hosting it. Until that changes, or until a neutral lab replicates the agentic-benchmark gains DeepSeek is claiming, the responsible read is this: a genuinely capable, sharply repriced coding model with one gold-standard independent result — and a benchmark chart that is still, for now, DeepSeek's word against nobody's check.