DeepSeek V4 Pro: Frontier Coding Performance at 7× Lower Cost Than Claude Opus
Agentic AIDeepSeek V4-Pro scores 80.6% on SWE-bench Verified — within 0.2 percentage points of Claude Opus 4.6. It costs $3.48 per million output tokens. Claude Opus charges $25. The 7× price gap at near-identical coding benchmark performance is the defining number in the April 2026 agentic AI infrastructure conversation.
Context
Agentic coding workloads are token-hungry in a way that chat interfaces are not. A single coding agent session resolving a multi-step feature request — reading files, drafting code, running validation, iterating — can consume hundreds of thousands of tokens. At scale, with agents running in parallel across development teams, API pricing shifts from a line item to a fundamental architecture constraint.
Until now, the high-capability options for coding agents clustered around OpenAI and Anthropic’s frontier models at $15–$25 per million output tokens. DeepSeek V4-Pro changes that calculus with a 1.6-trillion-parameter model that trades within a fraction of a percent of the current frontier — at a price point competitive with mid-tier options.
The Detail
DeepSeek released two V4 SKUs simultaneously. V4-Pro carries 1.6 trillion parameters with a one-million-token context window, priced at $1.74 per million input tokens and $3.48 per million output tokens. V4-Flash runs at 284 billion parameters, priced at $0.14 input and $0.28 output — positioned for high-throughput, latency-sensitive tasks.
On SWE-bench Verified, the standard benchmark for real-world software engineering task completion, V4-Pro posts 80.6% — within 0.2 points of Claude Opus 4.6 and above GPT-5.4. On LiveCodeBench, V4-Pro leads the current field at 93.5, compared to Gemini at 91.7 and Claude at 88.8. MMLU scores 88.4%; the Humanities-X reasoning benchmark reaches 92.1%, matching or narrowly beating this quarter’s frontier models.
The training economics behind these numbers are as significant as the benchmarks. DeepSeek trained V4 on 16,000 Hopper-era GPUs at a total compute cost of $5.6 million — doubling training efficiency over V3. For context, that compute spend is well below what US frontier labs report for comparable capability levels. The model remains available via DeepSeek’s API and through partners, with the V4-Flash tier offering developers a path to AI code quality at token costs that make high-volume automation economically viable.
The So What
The architectural decision for agentic coding pipelines is no longer purely a capability question. V4-Pro’s benchmark position means teams can evaluate DeepSeek as a primary model — not a fallback — for coding agents, with pricing that changes the cost model for workloads running millions of tokens per day.
The 7× price gap compounds in agent pipelines: coding agents that orchestrate sub-agents, run parallel tasks, or iterate through multiple drafts amplify token consumption. At $3.48 versus $25 per million output tokens, the cost difference for a high-volume agentic workflow reaches orders of magnitude over a quarter. V4-Flash extends that argument further — at $0.28 per million output tokens, it enables agent patterns that were economically impractical at frontier pricing.
Photo by Markus Winkler on Pexels
Source: Simon Willison's Weblog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →