On 12 August 2026, DeepSeek moved deepseek-v4-pro out of preview. The served build is now DeepSeek-V4-Pro-0813, a 1.6T-parameter mixture-of-experts model with a 1M-token context window. As with the July Flash refresh, the model ID did not change: you still request deepseek-v4-pro and are handed the GA build, so existing agents migrated with no config edit.
The Benchmark Jump
DeepSeek's published comparisons against the preview build are lopsided toward agent and tool work:
- Terminal Bench 2.1: 72.1 to 87.9
- DeepSWE: 12.8 to 62.7
- CyberGym: 52.7 to 83.3
- DSBench-Hard: 31.1 to 67.2
- NL2Repo: up 23.0 points to 61.5
- Automation Bench: 12.8 to 31.8
These are DeepSeek's own numbers on agent-flavoured benchmarks rather than independent evaluations, and a DeepSWE jump from 12.8 to 62.7 says as much about how weak the preview was on that test as about how strong the GA build is. The model also remains text-only, with no image or video input, and it is not the fastest way to get an answer — long agent runs on V4 Pro finish noticeably behind the frontier subscription tiers.
It Reverses Our July Advice
Between 31 July and 12 August, the cheap V4-Flash-0731 build genuinely led the V4 Pro preview on several agent benchmarks — 82.7 on Terminal Bench 2.1 against 72.1, and 54.4 on DeepSWE against 12.8. The GA Pro build takes that lead back on both. Flash is still roughly a third of Pro's output cost and carries five times the concurrency, so it remains the right default for high-volume chat and cheap tool loops; it is no longer the stronger model.
Pricing, and the Warning Attached to It
Direct API pricing is unchanged: $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens. For scale, that is roughly an order of magnitude under frontier flagship rates, and the cache-hit rate is where the gap gets absurd for repeated-context agent loops. DeepSeek has also stated on its own pricing page that an overall API price rise is coming, described as significant — so anyone building a cost model on these numbers should treat them as a snapshot rather than a floor.
DeepSeek V4 Pro is in the model picker on OpenClaw Launch, and any instance already running it is on the 0813 build. Setup steps and the full comparison against Flash are in the DeepSeek V4 Pro guide.