Daily AI CatchupDeepSeek V4-Pro Undercuts OpenAI by 8x on Coding Agent Benchmarks
DeepSeek released V4-Pro-0813, priced at $0.87 per 1 million output tokens, significantly cheaper than competing models. It successfully runs OpenAI's coding agent benchmarks at roughly 1/8th the cost of comparable alternatives. DeepSeek also open-sourced Harness, the agent framework it used for benchmarking, giving developers direct access to their evaluation methodology.
Why it matters
💻 Developer · DeepSeek Harness is now open-source and reproducible. You can audit their benchmarks directly and use the same framework to evaluate your own agent pipelines.
📦 Product · At $0.87 per million output tokens, the math on building agent-heavy features shifts dramatically. Cost per user drops by an order of magnitude, enabling previously uneconomical use cases.
🎨 Design · Cheaper inference means faster iteration on experimental features. You can prototype agent-assisted UI flows without running into budget constraints.
📈 Business · Price pressure from DeepSeek forces the entire market lower. If you're locked into expensive provider contracts, this is leverage in renegotiation. New ventures can now build profitable agent products at lower burn rates.
🤔 Just Curious · This is the clearest evidence yet that the AI cost curve is still collapsing. Open-sourced benchmarking also breaks the 'trust us on our numbers' dynamic that dominated earlier model releases.
Try this: Pull DeepSeek Harness and run your current agent pipeline against both DeepSeek V4-Pro and your existing provider. Calculate actual cost-per-task. The gap is often larger than you'd expect from published per-token rates.
Sources: DeepSeek V4-Pro Goes Live and Runs OpenAI's Own Coding Agent 8x Cheaper, DeepSeek Harness Opens the Agent Framework It Used to Benchmark Its Own Models