跪拜 Guibai
← All articles
Artificial Intelligence · LLM

DeepSeek V4-Flash Ships a 13B-Active-Param Agent That Beats Its Own Pro Preview

By 土豆1250 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 13B-active-parameter model now handles full-repo code analysis, multi-tool agent workflows, and terminal automation at a fraction of flagship cost. Native Codex compatibility turns OpenAI's monetization entry point into a customer-acquisition channel for DeepSeek, and the 98% cache discount resets the unit economics for any team running high-volume agent workloads.

Summary

A single changelog line on July 31, 2026 dropped a lightweight model that upends the industry's cost-versus-capability trade-off. V4-Flash keeps the same 284B-total / 13B-active-parameter architecture as its preview but retrained only the post-training stage, pushing Agent scores past the heavier V4-Pro-Preview on all nine reported benchmarks. Independent evaluators back the claim: Artificial Analysis scores it 50, one point behind GPT-5.6 Luna, while Frontend Code Arena ranks it third in the open category.

Native Responses API support means any Codex user can swap `base_url` and model name and start calling a cheaper Chinese model with zero client changes. The price table didn't budge, but a roughly 98% cache-hit discount makes per-task costs about 60% lower than OpenAI's already-discounted Luna. Peak/off-peak pricing is also teed up, doubling rates during Beijing business hours.

The old `deepseek-chat` and `deepseek-reasoner` endpoints went dark a week before the announcement, meaning all traffic was already flowing through V4-Flash under the preview label. A V4-Pro official release is promised "as soon as possible," with Responses API support expected in early August.

Takeaways
— V4-Flash's architecture and parameter count are unchanged from the preview; only post-training was redone, yet it beats V4-Pro-Preview on all nine reported Agent benchmarks.
— Artificial Analysis gives it an Intelligence Index of 50, one point behind GPT-5.6 Luna, while Frontend Code Arena ranks it third in the open category at 1586 points.
— Native Responses API support lets Codex users switch to DeepSeek by changing `base_url` and model name with no client modifications.
— Input pricing stays at $0.14/M tokens (cache miss) and $0.0028/M tokens (cache hit), with output at $0.28/M tokens; the ~98% cache-hit discount drives per-task costs roughly 60% below GPT-5.6 Luna.
— Peak/off-peak pricing is previewed: 2x rates during Beijing daytime hours (9:00–12:00, 14:00–18:00), pushing batch workloads to nights and weekends.
— The old `deepseek-chat` and `deepseek-reasoner` endpoints were silently decommissioned on July 24; all traffic had already been routed to V4-Flash for over a week before the official announcement.
— V4-Pro official version is slated for release "as soon as possible," with Responses API support expected in early August.
— The current Responses API implementation is stateless: `previous_response_id` is unsupported and `store` is always false, so conversation history must be managed client-side.
Conclusions

Post-training alone closed the gap between a lightweight model and a heavier preview, suggesting the base model was under-exploited and that inference-time behavior is now the primary lever for capability gains.

Native Codex compatibility is a distribution play, not a technical integration. It lets DeepSeek acquire users inside OpenAI's own ecosystem without asking them to learn new tooling.

A 98% cache-hit discount changes the cost calculus more than a headline price cut. In agent loops where system prompts and tool definitions repeat every turn, effective cost can drop by orders of magnitude.

Silently routing all traffic through the new model for a week before announcing it is a deployment pattern that uses live production traffic as the final QA stage, reducing rollout risk without a formal canary infrastructure.

The nine benchmarks are almost exclusively Agent and code tasks. General chat and writing capabilities are unmeasured, so the model's strength is narrow but deep in the workflows that drive API revenue.

Concepts & terms
Post-training
A phase after initial model training that fine-tunes behavior—how a model follows instructions, uses tools, retries on failure, and decides when to stop—without changing the underlying architecture or parameter count.
Responses API
OpenAI's newer interface specification that carries structured information like reasoning traces and tool calls, used by Codex, the Agent SDK, and automation tooling as the standard protocol for agentic workflows.
Cache-hit discount
A price reduction applied when input tokens match previously cached content. In agent loops, system prompts and tool definitions repeat every turn; a high cache-hit rate can slash effective cost by 50x or more compared to uncached input.
Active parameters
The subset of a model's total parameters actually used during a single forward pass, often via a Mixture-of-Experts architecture. V4-Flash has 284B total parameters but only activates 13B per token, keeping compute low while retaining a large knowledge base.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗