DeepSeek V4 Pro became generally available on 13 August 2026, when the Chinese AI developer DeepSeek moved its top model out of preview on its app, its website and its API. The release came with something unusual for an AI service: time-of-day pricing. From 16:00 UTC on 16 August, software that calls the model outside DeepSeek's peak hours pays half the peak rate. For Australian businesses there is a catch, because those peak hours fall in the middle of the Australian working day.
What happened
DeepSeek first released its V4 models as a preview on 24 April. In the V4 Pro release note of 13 August, the company announced the finished version and listed three changes:
- Agent upgrades. DeepSeek says the model is considerably better at agent work, meaning multi-step tasks where the model uses tools such as a code editor or a browser instead of only answering a question.
- Adjustable reasoning effort. V4 Pro and the smaller V4 Flash now offer three settings: low for simple tasks, high for everyday agent workflows and max for complex ones.
- Easier connection to existing tools. The API now natively supports the OpenAI Responses format, a common way for software to talk to AI models, with a one-click setup for the Codex coding tool.
In the app and on the web, V4 Pro sits behind a setting called Expert Mode. For developers the model name is unchanged, so software already using the preview moved to the new version without edits.
DeepSeek also published the model itself the same day. The DeepSeek-V4-Pro-0813 model card on Hugging Face lists 1.7 trillion parameters and the MIT licence, a short and permissive open-source licence. The rollout was not tidy. Simon Willison noted on his weblog that the model appeared on the API before DeepSeek had any obvious announcement page, and that, as far as he could tell, the benchmark figures were first released to DeepSeek's official WeChat group and reached other forums as copies.
Key details
The new API prices, from the table in DeepSeek's release note, are in US dollars per million tokens (a token is a word or part of a word):
| Model and period | Input (cached) | Input (not cached) | Output |
|---|---|---|---|
| V4 Pro, off-peak | $0.022 | $0.66 | $1.98 |
| V4 Pro, peak | $0.044 | $1.32 | $3.96 |
| V4 Flash, off-peak | $0.007 | $0.22 | $0.66 |
| V4 Flash, peak | $0.014 | $0.44 | $1.32 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and every other hour is off-peak. In Australian time that works out as follows:
- Sydney, Melbourne and Brisbane (AEST): 11am to 2pm and 4pm to 8pm.
- Perth (AWST): 9am to midday and 2pm to 6pm.
- When daylight saving starts in the south-eastern states, their windows move an hour later.
The benchmark table in the release note comes from DeepSeek's own testing and has not been independently confirmed. It shows a large jump from the April preview, with Claude Fable 5 (which DeepSeek lists as tested "w/ fallback") still ahead on these three tests:
| Test (DeepSeek's figures) | V4 Pro preview | V4 Pro, 13 August | Fable 5 (w/ fallback) |
|---|---|---|---|
| Terminal Bench 2.1 (command-line tasks) | 72.1 | 87.9 | 88.0 |
| DeepSWE (software engineering) | 12.8 | 62.7 | 70.0 |
| HLE without tools (hard exam questions) | 37.7 | 42.7 | 53.3 |
Why it matters
Time-of-day pricing changes how an AI budget works. Electricity has long been cheaper when demand is low. DeepSeek is applying the same idea across its V4 models, and saying so plainly: the release note describes the off-peak rate as a way to allow "more flexible workload scheduling". Work that does not need an instant answer, such as summarising yesterday's documents or checking a large batch of records, can be timed for the cheaper hours.
The new rates are an increase on the old ones. On 14 August, an archived copy of DeepSeek's pricing page still listed V4 Pro at US$0.435 per million input tokens that are not cached and US$0.87 per million output tokens, with the peak and off-peak table shown as the prices to come. The new off-peak output rate of US$1.98 is more than double that figure, and the peak rate of US$3.96 is more than four times it. "Half price off-peak" describes the gap between the two new rates, not a saving on what developers were paying before.
Open weights widen the choice of host. Because the model is published under the MIT licence, other hosting companies can run it on their own servers. A business that wants the model but not DeepSeek's own service may be able to get it from a provider in a different country, at that provider's price.
Data location is the real decision. DeepSeek's privacy policy, last updated on 10 February 2026, says the company collects, processes and stores personal data in the People's Republic of China. It also says personal data may be used to train and improve its models, and it lists, among the privacy rights that may be available to a user, a right to opt out of that use. For developers, the policy adds that information collected from the end users of software built on DeepSeek is governed by the developer's own privacy rules. In other words, if your software sends customer details to DeepSeek, explaining and protecting that is your job.
What this means for businesses
Few small businesses will sign up for DeepSeek's API themselves. The more likely path is that a developer, an agency or a software product you use chooses it because of the price. That makes this a supplier question as much as a technology one.
Under the Privacy Act, sending personal information to an overseas service carries obligations. The OAIC's guidance on Australian Privacy Principle 8 says an organisation must take reasonable steps to ensure the overseas recipient does not breach the principles, and it generally stays accountable if the recipient does.
A short checklist:
- Ask your developer or software supplier which AI models your systems call, and whether each is reached through its maker or through another host.
- Decide in writing which kinds of information may be sent to which services. Public web content and your own marketing drafts are a different category from client files, payroll or health details.
- If you do use the API, schedule bulk jobs outside the peak windows. On Australian Eastern Standard Time that means before 11am, between 2pm and 4pm, or after 8pm.
- Match the reasoning effort to the task. DeepSeek recommends the low setting for simple tasks.
- Set a spending limit and review usage monthly, as you would for any metered service.
- Test on your own real tasks before relying on a vendor's benchmark table.
We help businesses connect and review the systems they depend on. Our business systems and integrations work covers how your software exchanges data with outside services, our cyber security service covers access and data-handling reviews, and IT consulting can help you weigh up suppliers. You can request a quote at any time.
Key takeaways
- DeepSeek V4 Pro left preview on 13 August 2026 on the app, the web and the API, and its weights were published under the MIT licence the same day.
- From 16 August, API use costs half as much off-peak as at peak, and both new rates for V4 Pro are higher than the prices listed before the change.
- Peak hours cover late morning to early afternoon and late afternoon to evening on Australia's east coast.
- DeepSeek's own tests show big gains over the preview, with Claude Fable 5 still ahead on most of the tests where both were scored.
- DeepSeek's service stores personal data in China, so decide what may be sent to it before anyone starts using it.
Frequently asked questions
What is DeepSeek V4 Pro?
It is the larger of DeepSeek's two V4 AI models. DeepSeek says the release version is much better than the preview at multi-step agent tasks. It is available in DeepSeek's app and website through Expert Mode and to software through an API.
When are DeepSeek's peak hours in Australian time?
The peak windows are 01:00 to 04:00 and 06:00 to 10:00 UTC. On Australian Eastern Standard Time that is 11am to 2pm and 4pm to 8pm, and in Perth it is 9am to midday and 2pm to 6pm.
Is DeepSeek V4 Pro open source?
The model's weights are published on Hugging Face under the MIT licence, so other organisations can download, run and host it. Using DeepSeek's own app or API is a separate hosted service with its own terms.
Where does DeepSeek store data?
DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China. A third-party host running the open weights would have its own location and terms.
What do the reasoning effort settings do?
In the words of the model card, they control how much deliberation the model spends before answering. DeepSeek recommends low for simple tasks, high for everyday agent workflows and max for complex ones.