跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

DeepSeek V4 Flash Ships a Quiet, Brutal Agent Upgrade That Undercuts Its Own Pro Model

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 13B-activated-parameter model now outperforms its own Pro variant on agent tasks while costing a fraction of frontier alternatives. For teams running large-scale tool-calling or sub-agent pipelines, the price-performance ratio shifts the default away from more expensive endpoints.

Summary

DeepSeek quietly pushed the official V4 Flash model, and the benchmark gains are not incremental. DeepSWE shot from 7.3 to 54.4, CyberGym added 38 points, and Terminal-Bench 2.1 climbed to 82.7. Every one of the nine agent benchmarks improved, and the Flash variant now leads the larger V4 Pro Preview on all of them — despite using the same 284B-total, 13B-activated parameter architecture. Only post-training changed.

Against external models, V4 Flash enters the top tier without taking the crown. GPT-5.6 Sol and Claude Opus 5 still hold leads on DeepSWE and Terminal-Bench, but Flash closes the gap on Toolathlon-Verified and AutomationBench to single digits. For long-chain engineering tasks, the frontier models remain the safe bet; for high-volume code review, repo analysis, and sub-agent orchestration, Flash becomes the cost play.

Pricing stayed flat at $0.14 input, $0.0028 cache-hit input, and $0.28 output per million tokens. That makes output roughly one-quarter the cost of OpenAI's cheapest GPT-5.6 Luna, with better agent scores across the board.

Takeaways
DeepSWE score rose from 7.3 to 54.4, a 47.1-point gain over the Flash Preview.
CyberGym gained 38 points, DSBench-Hard gained 33.8, and DSBench-FullStack gained 31.7.
Terminal-Bench 2.1 reached 82.7, up 20.9 points from the Preview.
V4 Flash now leads V4 Pro Preview on all nine reported agent benchmarks.
The architecture stayed at 284B total parameters with 13B activated per token; only post-training changed.
GPT-5.6 Sol still leads Flash by 18.6 points on DeepSWE and by 5–6 points on Terminal-Bench 2.1.
Flash's output price of $0.28 per million tokens is roughly one-quarter of GPT-5.6 Luna's $1.20.
Cache-hit input at $0.0028 is 86% cheaper than Luna's $0.02.
Conclusions

Post-training alone delivered gains that normally require a model-size jump, which suggests agent-specific RL fine-tuning is now a higher-leverage investment than scaling parameters.

Flash beating Pro Preview across the board while sharing the same architecture implies the Pro variant was undertrained on agent tasks, not architecturally limited.

DeepSeek's decision to ship this without a formal announcement signals confidence that the numbers speak for themselves — and possibly that a V4 Pro official release is close enough to make Flash a stepping stone rather than the headline.

At Flash's price point, the cost of running agentic workflows at scale drops below the threshold where many teams would default to a frontier model for non-critical tasks.

Concepts & terms
DeepSWE
A benchmark that evaluates a model's ability to resolve real-world software engineering tasks, such as fixing bugs or implementing features in a codebase.
Terminal-Bench 2.1
A benchmark measuring how well an AI model can operate a command-line terminal to complete multi-step system administration and development tasks.
Toolathlon-Verified
A benchmark suite that tests a model's ability to select and correctly use external tools and APIs across diverse, verified task scenarios.
Post-training
The phase after initial pre-training where a model is fine-tuned using techniques like reinforcement learning or supervised fine-tuning to improve specific capabilities without changing the base architecture or parameter count.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗