DeepSeek GA'd DeepSeek-V4-Pro-0813 on August 13 after a four-month preview phase, optimizing for agentic AI tasks: tool use, code execution, and multi-step workflows. The model achieved 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE (software engineering), and 61.5 on NL2Repo. It supports 1M-token context with up to 384K-token outputs, dual thinking and non-thinking modes, and three configurable effort levels (low/high/max). API access is live; prices scale from flat-rate $0.87/M output tokens to peak-hour $3.96/M (Aug 16), with off-peak 50% discounts.
The general availability release marks a strategic shift from the April preview, which was outperformed by DeepSeek's own cheaper V4-Flash model in some independent tests—an unusual outcome that created early momentum for Flash. With the 0813 GA, DeepSeek is now pricing V4-Pro at 14× the output cost of Flash ($3.96 vs $0.28) while claiming substantive agent improvements in production workloads. The company also announced OpenAI API format support, Codex integration, and native Responses API compatibility, positioning the model for enterprise agentic workflows.
For builders, the relevance is cost-vs-capability routing: Together AI benchmarked the cascade (run Pro first, escalate to GPT-5.6 Sol on test failure) at $3.35 per solved task—outperforming Sol alone (72.7% pass@1) at less than half the per-task price. At 1/35th the unit cost of GPT-5.6 Sol and with strong agentic agent coverage on open benchmarks (80.6% on SWE-Bench Verified), DeepSeek V4-Pro is forcing a fundamental repricing of coding-agent economics.