DeepSeek Raises API Prices as V4-Pro Gets Sharper

Six times, and twelve times at peak hours. That is how much the price of a cache hit is going up on DeepSeek's API from August 16.

The company shipped an upgraded V4-Pro on August 13 and announced new pricing in the same breath. The model got meaningfully better. It also stopped being the cheapest thing on the shelf by quite the same margin.

The benchmark movement

V4-Pro-0813 leaves preview as a stable build, and the coding numbers moved a long way:

• Terminal Bench 2.1: 72.1 to 87.9
• DeepSWE: 12.8 to 62.7
• Artificial Analysis intelligence index: 45 to 53

The DeepSWE result is the one worth pausing on. Roughly a fivefold improvement on a benchmark built around real software repositories suggests the model became genuinely usable for repo-scale work rather than snippet generation. Even so, 53 on the intelligence index still trails Claude Opus 5 at 63. The distance narrowed; it did not disappear.

Two additions on the engineering side: the 1 million token context window carries over unchanged, and the model now natively supports OpenAI's Responses API along with Codex integration. Reasoning effort can be set to low, high or max.

Where the increase actually lands

Standard token pricing rises in a predictable way. Off-peak input goes from $0.435 to $0.66 per million, output from $0.87 to $1.98. During peak windows both double again, to $1.32 and $3.96. Peak is defined against Chinese working hours, 01:00 to 04:00 and 06:00 to 10:00 UTC.

Cache hits are the outlier. They move from $0.003625 per million tokens to $0.022, and to $0.044 at peak.

That line matters more than its size suggests, because cached tokens are where high-volume applications live. A support assistant that resends the same system instructions, product catalog or policy document on every request is serving most of its traffic from cache by design. Teams that optimized their unit economics did it precisely there. DeepSeek is partially reversing the cuts it made in May 2025, and the timing coincides with preparation for a capital raise.

Harness arrives under an MIT license

The second half of the announcement pushes the other way. DeepSeek open-sourced its agent software, Harness, as a v0.1 developer preview under the MIT license.

Harness turns a language model into an autonomous agent through modular plugins. It logs sessions, lets runs be paused and resumed, replays past runs step by step, and ships with a local web interface.

The license choice is not a footnote. MIT is about as permissive as open source gets, with no meaningful obstacle to commercial use, which means you can run the software on your own infrastructure against whichever model you prefer. An agent framework that does not bind you to the vendor who wrote it is a useful thing to have appear on the market.

Reading the two halves together

Taken as one announcement, the message is that DeepSeek is repositioning. It is climbing on capability, normalizing on price, and investing in tooling that is deliberately not locked to its own models.

For anyone budgeting around a low-cost provider, the lesson outlives this particular price list. A cost plan anchored to one vendor's pricing is an assumption with an expiry date you do not control. It belongs in a forecast that gets revisited quarterly, not in a spreadsheet cell treated as fixed.

Two practical steps. Measure your real token consumption instead of estimating it, because most teams are estimating. Then find out what share of it is served from cache. Without those two numbers, no pricing announcement can be translated into an impact on your business, and the question of whether to switch providers stays an argument. With them, it becomes arithmetic.

Sources: The Decoder, DeepSeek