What DeepSeek sells
DeepSeek is the Chinese company selling exactly two models today: V4-Flash at $0.14 in and $0.28 out, and V4-Pro at three times that price. The part that rarely gets written down is that DeepSeek own change log says the Flash results run far ahead of V4-Pro-Preview.
- $0.28 output
- one million token window
- a price rise announced in the docs
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Model families
One model line with two price steps, and the gap between them is threefold.
Versions
| Versions | Released | Status | Model families |
|---|---|---|---|
| DeepSeek V4 Flash | current | DeepSeek V4 |
The cheaper model leads on the maker own benchmarks
In the DeepSeek change log, the 31 July 2026 entry says V4-Flash shipped with significantly enhanced agent capability and results far exceeding V4-Pro-Preview. The three figures printed beside it are Terminal Bench 2.1 at 82.7, NL2Repo at 54.2 and Cybergym at 76.7. Architecture and model size did not change; it was re-post-trained.
That leaves you a simple decision. Make Flash the default. Pro costs three times more and the maker has not shown its advantage on the same page.
The number that makes DeepSeek cheap, and its condition
V4-Flash input carries two rates: $0.14 on a cache miss and $0.0028 on a hit. The gap is fifty-fold, and our tables always carry the higher rate, because no real workload hits cache every time. Even at $0.14 this is the cheapest row in our coding table.
One more difference gets little attention: the concurrency ceiling is 2,500 requests on Flash against 500 on Pro. The cheaper model also carries five times the parallel capacity, and for a service with bursty traffic that can matter more than the price does.
The warning printed on the pricing page
That same pricing page states a plan to raise overall API pricing in the near future, with a significant increase expected. It belongs on this page because if you have budgeted on $0.28 output, or priced your own product from it, your cost model has an expiry date. A vendor that says so in advance is behaving well, but it means today figure has no business inside a long contract.
Where it genuinely leads, and where it does not
The DeepSeek change log is the most transparent in this group: every version is dated, from R1 in January 2025 to V4 on 24 April 2026, and even the retirement of the old deepseek-chat and deepseek-reasoner ids on 24 July 2026 is written down. Moonshot and Z.ai publish nothing of the kind.
Against that, DeepSeek total in our coding table is 20 of 100 and all of it comes from the price column. In the biggest column, the WebDev leaderboard, it scores 1585 against 1692 for Opus 5. Cheapest means cheapest, not best.
What it is good at
- The cheapest model in our whole coding table, at $0.14 in and $0.28 out
- A one million token window with up to 384K of output, the highest output ceiling in this group
- A complete dated change log, including the date the old model ids were switched off
- The API speaks both the OpenAI and the Anthropic format, so Claude Code and GitHub Copilot connect with no code change
- A concurrency ceiling of 2,500 requests on the cheaper model
Where it falls short
- The pricing page states plainly that prices will rise significantly in the near future. Today figure is not something to write into a long contract.
- It scores 1585 on the WebDev leaderboard, a hundred and seven points behind Opus 5. Cheap does not substitute for capable.
- It publishes no supported-countries list, so our Iran column stays unchecked.
- The change log announces no open weights release for the V4 generation, and the home page links GitHub repositories without stating licence terms on the page itself.
- The entire catalogue is two text models. There is no image, video or audio model in the API documentation.
Our take
If you have volume, a small budget and no need for top-tier quality, DeepSeek is the serious cheap route, and the model to pick is Flash rather than Pro. Just do not fix a long budget to today price, because the maker has already said it is going up.
Questions people actually ask
Is V4-Flash or V4-Pro the better model
DeepSeek own change log says the agent results for Flash run far ahead of V4-Pro-Preview, while Flash costs a third of Pro and carries five times the concurrency ceiling. Until a fresher number is published for Pro, Flash is the reasonable default.
Will DeepSeek prices stay where they are
No, and the vendor says so. The pricing page announces a significant increase in the near future. If your business model leans on that number, cost the rise in now.
Does code written for Claude connect to DeepSeek
Yes. The documentation says the API accepts both the OpenAI and the Anthropic format, and that tools such as Claude Code and GitHub Copilot work without code changes. Output quality is a separate question, and the ranking table answers it.
Sources
- DeepSeek API pricing vendor source 12 August 2026
- DeepSeek change log vendor source 12 August 2026
- DeepSeek vendor source 12 August 2026