DeepSeek Ships V4 Pro – and Raises API Prices Sharply
DeepSeek moves V4-Pro into production, open-sources its agent framework – and raises API prices by up to 1,100 percent, effective August 16.
Illustration · AI-generated (AI IN LIFE)
At a glance
- V4-Pro-0813 is in production; the context window stays at one million tokens.
- API prices rise by up to 1,100 percent per Caixin, effective August 16, 2026.
- Cache hits get pricier: from $0.003625 to $0.022 per million tokens (off-peak).
- Benchmarks: Terminal Bench 2.1 up from 72.1 to 87.9; DeepSWE from 12.8 to 62.7.
- Agent framework „DeepSeek Harness v0.1“ ships under MIT license; 712 beta projects in three days.
DeepSeek has officially moved its flagship model V4-Pro (version 0813) from testing into production – and announced a hefty price increase at the same time. According to Caixin, some API prices rise by as much as 1,100 percent. The Decoder and Android Headlines also covered the move.
The model keeps its parameter count and one-million-token context window. New additions include native support for OpenAI's Responses API with Codex integration and adjustable reasoning effort levels („low“, „high“, „max“).
On agentic benchmarks, V4-Pro improves markedly: Terminal Bench 2.1 jumps from 72.1 to 87.9, DeepSWE from 12.8 to 62.7. On Artificial Analysis' Intelligence Index, however, the model scores 53 – still behind Claude Opus 5 (63), Kimi K3 (60) and Qwen 3.8 Max (58).
The new prices take effect August 16: DeepSeek introduces peak and off-peak rates based on Chinese business hours. Off-peak input rises from $0.435 to $0.66 per million tokens. Cache hits are hit hardest: from $0.003625 to $0.022 – especially relevant for agent workflows that repeatedly read the same files. Even so, DeepSeek remains far cheaper than Western frontier models.
In parallel, the company is opening up its agent software: „DeepSeek Harness v0.1“ launches under an MIT license as a modular framework with session tracking, run resumption and branching. The developer preview attracted 712 beta projects within three days, per the company.
FAQ
Why is DeepSeek raising prices?
The company is monetizing its flagship: peak/off-peak rates and pricier cache hits primarily target heavy agent workloads. DeepSeek still stays cheaper than Western frontier models.
What changes for agent developers?
Cache-hit pricing bites hardest, rising from $0.003625 to $0.022 – costly for agents re-reading identical files. In return, Harness v0.1 offers an open MIT-licensed framework.
How strong is V4-Pro comparatively?
It gains massively on agent benchmarks, but scores 53 on Artificial Analysis' Intelligence Index – behind Claude Opus 5, Kimi K3 and Qwen 3.8 Max.
