The technical map of DeepSeek infrastructure
DeepSeek is a Chinese company that ships its model infrastructure as two units: V4-Flash, rated at $0.14 for input and $0.28 for output, and V4-Pro, at three times that rate. A detail rarely surfaced elsewhere: DeepSeek own change log records that the cheaper unit results have moved past the more expensive unit preview.
- output rate: $0.28
- window capacity: one million tokens
- a price rise warning is on record
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Model families
One production line, with two pricing tiers three times apart.
Versions
| Versions | Deployed | Status | Model families |
|---|---|---|---|
| DeepSeek V4 Flash Vision | preview | DeepSeek V4 | |
| DeepSeek V4 Flash | current | DeepSeek V4 | |
| DeepSeek V4 Pro | current | DeepSeek V4 |
The cheaper unit scores higher on the maker own benchmarks
In the DeepSeek change log, the entry dated 31 July 2026 reports that V4-Flash shipped with substantially upgraded agentic capacity and that its results run "clearly beyond V4-Pro-Preview." Three metrics are logged alongside that claim: Terminal Bench 2.1 at 82.7, NL2Repo at 54.2 and Cybergym at 76.7. Architecture and model size were left unchanged; only the post-training pass was repeated.
The operational consequence is a simple rule: set Flash as default. Pro carries three times the cost without the maker demonstrating its advantage in the same documentation.
The parameter that lowers cost, and the condition attached to it
V4-Flash input runs on two operating rates: $0.14 on a cache miss and $0.0028 on a cache hit. The gap between the two rates is fiftyfold, and our tables always log the higher rate, since no real workload clears cache on every call. Even at the $0.14 rate, this remains the lowest-priced row in our entire coding table.
A second parameter draws less attention: the concurrency ceiling on V4-Flash is 2,500 requests against 500 on Pro. The cheaper unit therefore carries five times the parallel processing capacity, and for a service with bursty traffic load, this parameter can matter more than the rate itself.
The warning logged on the pricing page
The same pricing page states that a plan is in motion to raise overall API service rates in the near term, with a "significant" increase projected. We surface this clause because if your cost model runs on the $0.28 output rate, that parameter carries an expiry date. Documenting the change in advance is sound practice on the maker part, but it means today rate should not be locked into a long-term contract.
Where it genuinely leads, and where it does not
The DeepSeek change log is the most transparent document in this group: every version is timestamped, from R1 in January 2025 through V4 on 24 April 2026, down to the retirement date of the legacy deepseek-chat and deepseek-reasoner identifiers on 24 July 2026. Moonshot and Z.ai do not document at this level.
Against that, DeepSeek total score in our coding table is 38.1 of 100, of which twenty points come from the price column. On the largest single metric, the WebDev leaderboard, it scores 1585 against 1692 for Opus 5. Lowest cost is not equivalent to highest capability.
Documented strengths
- The lowest-cost unit in our entire coding table: $0.14 input, $0.28 output
- A one-million-token window with output capacity up to 384K tokens, the highest output ceiling in this group
- A complete, dated change log, including the retirement date of legacy identifiers
- The interface is compatible with two formats, OpenAI and Anthropic, so tools such as Claude Code and GitHub Copilot connect with no code change
- A concurrency ceiling of 2,500 requests on the cheaper unit
Known weaknesses and limits
- The pricing page formally states rates will rise significantly in the near term. Today rate is not a stable parameter for a long-term contract.
- It scores 1585 on the WebDev leaderboard, a hundred and seven points behind Opus 5. A low rate does not substitute for processing capability.
- No supported-countries list is published, so the Iran access column in our system is logged as unverified.
- The change log announces no open-weight release for the V4 generation, and the home page links GitHub repositories without documenting licence terms on that same page.
- The entire catalogue consists of two text models. No image, video, or audio unit is logged in the API documentation.
Technical verdict
For high-volume workloads on a constrained budget that do not depend on top-tier quality, DeepSeek is the most serious low-cost route, and the unit to specify is Flash, not Pro. Just do not lock a long-term budget to today rate, since the maker has already logged the coming increase.
Frequently raised questions, with documented answers
Does V4-Flash outperform V4-Pro
DeepSeek own change log records that Flash agentic results run ahead of V4-Pro-Preview, while Flash carries a third of the Pro rate and five times its concurrency ceiling. Until a fresher figure is published for Pro, Flash is the reasonable default.
Will DeepSeek rates stay fixed
No, and the maker has documented this itself. The pricing page logs a significant increase for the near term. If your operating model depends on this rate, factor the increase scenario into your calculations now.
Does code written for Claude run on DeepSeek
Yes. The documentation states the interface accepts both the OpenAI and the Anthropic format, and that tools such as Claude Code and GitHub Copilot function with no code change. Output quality is a separate question the ranking table addresses.
How much this tool is searched for
"deepseek" is searched 1,220,000 times a month in the United States and 22,200 times in Saudi Arabia. That is a 55x gap between two markets this site publishes in.
| Market | Term entered | Monthly search volume | Click cost | Ranking difficulty | User goal |
|---|---|---|---|---|---|
| Turkey | deepseek | 27,100 | $0.18 | 96 | Navigational |
| Saudi Arabia | deepseek | 22,200 | $0.21 | 87 | Navigational |
| United States | deepseek | 1,220,000 | $1.35 | 100 | Navigational |
Data from Semrush, TR, SA, US databases, retrieved 27 August 2026. Search volumes decay over time; read anything more than three months old as a trend line, not a fixed figure.