DeepSeek V4-Pro Hits GA and Triples Its Price

DeepSeek moved V4-Pro to general availability on 13 August 2026 across its app, web interface and API, under the model ID deepseek-v4-pro. Three days later, on 16 August at 16:00 UTC, the entire V4 family switched to peak/off-peak billing.

The new rates

Per million tokens, cache miss:

Cache-hit input on V4-Pro is $0.044 per million at peak and $0.022 off-peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; every other hour is off-peak at half the peak rate.

Correction to our own coverage: this page and our comparison guide previously listed V4-Pro at $0.435 input / $0.87 output per million tokens. That was accurate for the preview period and is now wrong - peak input is roughly three times higher. We have updated the model page, the specification table and the "cheapest API" claims in our top AI models comparison accordingly.

For context on what the increase does to the competitive picture: GPT-5.6 Sol is $5 input / $30 output per million. V4-Pro at peak is about a quarter of Sol's input rate rather than the tenth it used to be. It remains the lowest-priced model on our roster, but the margin is much narrower - GLM-5.2 sits at $1.4 / $4.4, only fractionally above V4-Pro's peak rate.

What GA changed technically

DeepSeek describes the GA build as substantially stronger at agent work - tool use, code execution, multi-step workflows - with the largest gains in production environments. It adds native Responses API support adapted for Codex integration, and three thinking-effort levels (low, high, max). The context window is 1M tokens with up to 384K tokens of output, in either thinking or non-thinking mode.

Vendor-reported benchmark figures released with the GA build: 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo.

A second model, deepseek-v4-flash, entered public beta on 31 July 2026. DeepSeek reports it beating V4-Pro-Preview on benchmarks while costing a third as much.