DeepSeek V4 Pro leaves preview: 1.6T flagship, 80.6% on SWE-bench
The staged rollout is done: the flagship is one API call at $0.435/$0.87 per Mtok, reportedly beats Opus 4.8 at agentic coding — with a price hike flagged.
Copy markdown
The 0813 build is the GA — and it's live
DeepSeek-V4-Pro-0813 stepped out of preview on Aug 12; call it at the `deepseek-v4-pro` endpoint, or via Together AI, DeepInfra and Nano-GPT. It's a 1.6T-parameter MoE (~49B active per token), 1M-token context, up to 384K output tokens.
The price is the headline: $0.435 in, $0.87 out
Per million tokens, with cache-hit input at $0.003625 and a 500-request concurrency cap. A 1.6T-class model now fits a hobby-project budget — a fraction of frontier closed-model rates.
Reportedly beats Opus 4.8 on agentic coding
DeepSeek reports ~80.6% on SWE-bench Verified and wins over Claude Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE and AutomationBench. Independent 0813 numbers aren't out yet, so treat these as vendor claims until verified.
A price hike is already flagged — lock nothing in
DeepSeek says a significant increase is planned but hasn't published new rates or a date. Build against today's pricing, but don't wire long-term cost forecasts to it.
Flash already drops into Codex; the Harness didn't ship
V4-Flash-0731 ($0.14/$0.28 per Mtok) has native Responses API support, so you can wire DeepSeek into the Codex CLI today. The promised DeepSeek Harness coding tool is still pending.
Open weights, with an asterisk
The V4 series shipped under MIT on Hugging Face back in April, but the served 0731/0813 builds are later post-training updates whose exact weights may differ from the published checkpoint. Self-hosters get the family, not necessarily this API model.
Elsewhere: Grok 4.6 lands the same week
xAI shipped Grok 4.6 on Aug 12 for long-running agents and coding at $2/$6 per Mtok — capable but roughly 3x DeepSeek's price. Together the two releases frame a clean cost-vs-capability call for agent builders.