Open-weight models
DeepSeek's V4-Pro reaches general availability, then prices jump up to 1,100%
The brief: a quiet pricing-page edit on 12 August, then peak/off-peak billing from 16 August.
The answer
DeepSeek's V4-Pro reached general availability on 12 August, then prices rose up to 1,100%.
DeepSeek's flagship model, V4-Pro, quietly reached general availability in mid-August 2026. Four days later, its price went up by as much as 1,100%.
DeepSeek-V4-Pro-0813 is now generally available, out of the preview it entered in April.
What happened
On 12 August, DeepSeek changed one line on its own API pricing page, listing DeepSeek-V4-Pro-0813 behind the existing deepseek-v4-pro endpoint. There was no blog post and no changelog entry. The preview it replaced had run since 24 April. Some outlets instead date general availability to 13 August, when the change was first widely reported.
The money
Launch pricing was $0.435 per million input tokens, $0.87 output, $0.003625 for a cache hit. From 16:00 UTC on 16 August, DeepSeek switched to peak/off-peak billing.
| Tier | Input | Output |
|---|---|---|
| Launch (to 16 Aug) | $0.435 | $0.87 |
| Off-peak | $0.66 | $1.98 |
| Peak | $1.32 | $3.96 |
Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday
Off-peak is billed as 50% cheaper than peak. It is still costlier than the flat rate that applied before 16 August, at every hour of the day.
The catch
DeepSeek's own model card claims big gains 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. None of those figures has independent replication. A neutral Terminus 2 harness run scored the Terminal-Bench figure at 54.68% — 33 points lower.
One benchmark is independently confirmed: SWE-bench Verified, where Vals.ai placed V4-Pro second of 82 models at 96.40%, behind only Claude Opus 5. On LiveBench's agentic-coding test, V4-Pro ranks last of seven frontier models.
The model
V4-Pro is a Mixture-of-Experts model: 1.6 trillion total parameters, 49 billion activated per token, a 1-million-token context window and hybrid attention. It has no vision support. No -0813-tagged weights have been published for self-hosting.
What's next
DeepSeek had signalled it would retire the deepseek-v4-pro endpoint on 14 September in favour of a newer V4.1 Flash architecture. It reversed that plan after demand held up, confirming continued service past that date.
DeepSeek has not said whether the 0813 checkpoint's weights will be published for self-hosting, or whether peak/off-peak billing will extend to other models in its lineup.
Frequently asked questions
When did DeepSeek-V4-Pro reach general availability?
How much more expensive did DeepSeek-V4-Pro get?
Is DeepSeek-V4-Pro's benchmark chart independently verified?
Is DeepSeek-V4-Pro still available?
Sources
- Models & Pricing — DeepSeek, 16 August 2026
- DeepSeek V4 Pro Launches GA, Then Hikes API Prices 1,100% — Enterprise DNA, 13 August 2026
- DeepSeek V4 Pro 0813 — The Quiet GA and What It Means — Floatboat, 14 August 2026
- DeepSeek-V4-Pro-0813 Benchmarks: How It Stacks Up Against Opus and Kimi K3 — MindStudio, 16 August 2026
- DeepSeek V4 Guide: Pro & Flash, GA + Pricing — Codersera, 13 August 2026