跪拜 Guibai
← Back to the summary

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

One-line positioning: This is an in-depth interpretation for developers, AI practitioners, and those who care about the large model industry—no hype, no bashing, breaking down the official changelog item by item, doing the math, and then boldly predicting DeepSeek's next move.

Why is this release worth putting down what you're doing? (Why)

On the afternoon of July 31, 2026, a line appeared in DeepSeek's API changelog: "DeepSeek-V4-Flash official version API is now in public beta."

No press conference, no livestream, no slides, not even a banner on the official website, and the official WeChat account was silent. But this single line pushed DeepSeek to the top of Zhihu's trending list that afternoon and evening, and the developer community exploded.

Why? Because this "official version" did something almost no one in the industry has managed in the past two years: a lightweight model called Flash beat its own Pro preview version into the ground. More subtly, just days earlier, OpenAI had slashed the price of GPT-5.6 Luna by 80%. In response, DeepSeek significantly boosted its Agent capabilities, kept the price unchanged, and casually pried open the door to OpenAI's Codex.

This isn't just tech news; it's a commercial battle over "pricing power" and the "Agent ecosystem entry point." Understanding this changelog means understanding the industry's direction for the next three months.

What exactly did the official changelog say? (What)

Same soup, different seasoning: Architecture unchanged, only the 'second half' was retrained

First, the most crucial sentence: The model structure and size of DeepSeek-V4-Flash-0731 are consistent with the Preview version; only post-training was redone.

In plain terms: the base model wasn't swapped, parameters weren't increased (total params 284B, active params only 13B). This time, only the "how to work" second half was retrained—how to receive tasks, how to use tools, how to retry on failure, and when to stop.

What does this mean? It means this upgrade wasn't a "new engine," but rather honing the "driving skills" to the extreme. Parameters didn't change, price didn't change, but the work got done better. In an industry obsessed with stacking compute to inflate parameters, this is a completely different path.

Nine benchmarks up across the board, Flash surpasses Pro Preview

The official changelog released 9 evaluation scores, all up, each surpassing its own V4-Pro-Preview:

Benchmark Score What it tests
Terminal Bench 2.1 82.7 Terminal operations, command-line task execution
Cybergym 76.7 Cybersecurity and offensive/defensive automation
Toolathlon (verified) 70.3 Tool-use capability
DSBench-FullStack 68.7 Full-stack development tasks (DeepSeek internal test set)
DSBench-Hard 59.6 High-difficulty development tasks (internal test set)
DeepSWE 54.4 Real GitHub repos: read issues, write patches, run tests
NL2Repo 54.2 Natural language to complete code repository generation
Agent Last Exam 25.2 Comprehensive agent examination
Automation Bench (Public) 25.1 Automated task execution

Item-by-item breakdown: What exactly do these 9 'Agent exams' test?

For readers who don't frequently read papers, these English names might look like gibberish. Don't worry, let's explain them one by one in plain language (first, a quick primer: a benchmark is a "standardized test"—think of it as a standardized exam where all models use the same paper and scoring criteria, making scores comparable):

Terminal Bench 2.1 (82.7) — 'Live practical exam' The closest to real work. It drops the AI into a real terminal environment (the black window developers type commands into daily) and gives it real tasks: modify config files, run scripts, install dependencies, fix bugs, dig through logs to locate problems. It doesn't test "knowing the answer," but "getting things done with your hands." A score of 82.7 means it can independently complete most terminal tasks—equivalent to an intern who can start working directly without hand-holding.

Cybergym (76.7) — 'Cybersecurity attack and defense training ground' Simulates real network attack and defense scenarios, examining the AI's reasoning and security operation capabilities in complex environments: identifying vulnerabilities, analyzing attacks, executing defensive actions. A score of 76.7 indicates strong performance in high-reasoning-intensity tasks like security, also reflecting the model's upper limit for complex reasoning.

Toolathlon (70.3) — 'Tool-use capability competition' "Athlon" originates from Greek, meaning "contest." This exam tests whether the AI can "use tools": given a goal, with a bunch of APIs, functions, and external tools laid out beside it, can it correctly select, correctly call, and process the returned results? Analogy: give a new hire a full IDE, browser, database client, and see if they can pick up the right tool at the right time—this is the core watershed between a "working Agent" and a "chat-only model."

DeepSWE (54.4) — 'Real software engineer exam' Drops the AI into a real GitHub repository, makes it read issues (user-submitted problem reports) by itself, locate relevant code, write patches, run tests until they pass, all without human intervention. This is one of the industry's universally recognized hardest software engineering benchmarks. According to community-circulated comparison data, Claude 3.5 Sonnet scores about 49% on the same metric, and GPT-4o about 38% (this cross-sectional data is not officially confirmed). A lightweight model called Flash reaching this score is already nipping at the heels of closed-source flagships.

NL2Repo (54.2) — 'Generate an entire code repository from one sentence' Short for Natural Language to Repository: input a natural language requirement (e.g., "make a blog system with user login"), directly output a complete, runnable code repository. It tests engineering organization ability from zero to one—not just writing a few functions, but scaffolding an entire project.

DSBench-FullStack / DSBench-Hard (68.7 / 59.6) — 'DeepSeek's own mock exams' Internally built test sets for full-stack development and high-difficulty tasks. Note: this is their own paper, graded by themselves—apply a mental discount when referencing.

Agent Last Exam (25.2) — 'Comprehensive final exam' A comprehensive exam for Agents, covering planning, reasoning, multi-step task execution, and other integrated abilities. A score of 25.2 looks low? Don't panic. This is an industry-renowned "hell-level" paper; all models generally hover in the 20s-30s. The absolute score cannot be compared horizontally with other benchmarks.

Automation Bench (25.1) — 'Automated pipeline execution' Tests the AI's ability to execute automated workflows: completing multi-step tasks step-by-step according to a process, handling intermediate exceptions. This is a core scenario for enterprise RPA (Robotic Process Automation) and ops automation.

Two important reminders:

  1. These scores are "official self-tests" run using their own unreleased Harness framework under specific parameter configurations; apply a discount when referencing.
  2. These 9 items are almost all Agent/code scenarios. General capabilities like chatting and writing may not be at the same level—don't treat this report card as a full-subject transcript.

Placed in the industry context: What level is this?

Looking comprehensively, the common direction of these numbers is very clear: DeepSeek is no longer making a "chat model," but is fully pivoting towards an "AI software engineer capable of completing work independently."

Opened a special door for Codex: Native Responses API support

A hidden bombshell beyond the benchmark scores is this unremarkable sentence: "V4-Flash natively supports the Responses API format and has been specifically adapted for Codex."

The Responses API is the next-generation interface specification launched by OpenAI last November; Codex, Agent SDK, and various automation tools all run through it. Previously, plugging a domestic model into Codex required a protocol conversion layer in between. The model's reasoning process, tool calls, and other structured information were easily flattened and lost during conversion.

Not anymore. Set base_url to https://api.deepseek.com, model name to deepseek-v4-flash, point directly at it and it works. Existing clients need zero line changes; unsupported parameters are silently ignored.

This isn't a technical detail; it's a business strategy. Codex is the core monetization entry point for OpenAI's GPT series—ChatGPT coding subscriptions, enterprise code automation, Cursor/Copilot backends all rely on it. DeepSeek plugging its "lightweight version" in using the same protocol is effectively telling all Codex users: you can use OpenAI's posture to call a cheaper Chinese model.

Price didn't rise a cent, but the cache discount is more aggressive

Capabilities jumped this much, what about the price? Unchanged, still the old price list:

Billing Item Price (USD / million tokens) Approx. RMB
Cache hit input $0.0028 Approx. 0.02 Yuan
Cache miss input $0.14 Approx. 0.95 Yuan
Output $0.28 Approx. 1.9 Yuan

The real killer move is the cache discount: DeepSeek provides approximately a 98% cache hit discount for its own API, while most industry players only offer around 90%.

Don't underestimate this few percentage point difference. In Agent scenarios, system prompts, tool definitions, and previous rounds of conversation history are sent repeatedly every round. If this portion hits the cache, the actual bill can differ by dozens of times compared to a miss. Calculated out, the per-task cost on DeepSeek's own API is about 60% lower than GPT-5.6 Luna, which already had its price cut by 80%.

Three things not advertised in the corner of the changelog

  1. Old APIs have been quietly taken offline: When the V4 preview version launched on April 24, the official announcement already warned that the two old interfaces deepseek-chat and deepseek-reasoner would cease maintenance within 3 months; on July 24 (from 15:59 UTC), the two old interfaces were officially deactivated, old requests directly fail, all traffic points to deepseek-v4-flash. That is to say, before the official version was announced, any DeepSeek function developers called was running V4-Flash behind the scenes—it ran quietly under the Preview label for over a week, digesting bugs in real traffic, before being "officially" announced.
  2. Peak/off-peak pricing preview: As early as June 29, when announcing the V4 official version launch plan, DeepSeek had already previewed the first-time introduction of a "peak/off-peak pricing" mechanism—peak hours (Beijing time 9:00–12:00, 14:00–18:00 daily) priced at 2x the normal rate, applicable to all billing items. The official note stated "specific times subject to formal notice." After this changelog release, this preview on the pricing page remains valid. For developers, running batch tasks off-peak can save half the cost.
  3. V4-Pro official version 'released as soon as possible': The last sentence of the changelog states the V4-Pro official version will be released as soon as possible. Official documentation shows Pro's Responses API support is expected to go live in early August.

Third-party perspective: What do independent evaluators say? (What +)

Official benchmark scores smell like "every peddler praises his own needles"? No worries, independent evaluators have results too.

Artificial Analysis (Intelligence Index AII): DeepSeek V4 Flash 0731 scored 50 points, 10 points higher than the V4 Flash released in April, 6 points higher than V4 Pro, and only 1 point behind OpenAI's current highest scorer, GPT-5.6 Luna (51 points). Even though OpenAI just cut Luna's price by 80% that same day, V4 Flash's per-task cost is still about 60% lower, the key factor being that 98% cache hit discount.

Arena (Human preference blind test): In the Frontend Code Arena, DeepSeek-V4-Flash-High reshaped the leaderboard with a score of 1586—ranked 7th overall, 3rd in the open category, a 154-point improvement over the Preview version, and a 121-point improvement over V4-Pro-Preview. Priced at $0.14/$0.28 per MToken, it was evaluated as "the best cost-performance in its class."

Independent verification from two dimensions points to the same conclusion: This isn't marketing fluff; Flash's Agent capability has genuinely risen.

How to use it? Developer integration posture (How)

Quick start

For developers already using the DeepSeek API, migration cost is nearly zero—model name stays deepseek-v4-flash, base_url unchanged. The only thing to do is re-run a round of regression testing:

base_url:  https://api.deepseek.com   (remains unchanged)
model:     deepseek-v4-flash          (old interface traffic has all been pointed here)

If you want to use Codex, just add the responses.create call according to the official Responses API guide.

A key reminder

The current Responses API is a stateless implementation: previous_response_id is not supported, the store parameter is constantly false, and you still need to carry the conversation history for each round yourself. Also, currently only Flash supports the Responses API; Pro will have to wait until early August.

How to calculate the cost

In Agent scenarios, system prompts, tool definitions, and previous rounds of history are sent repeatedly every round. If this portion can hit the cache, the actual bill can differ by dozens of times compared to a miss. So in this price list, the line most worth watching is the 0.02 Yuan one—designing the prompt structure well and keeping the cacheable parts stable is the only correct way to save money.

What does this release mean for the industry?

1. The formula 'lightweight = cheap + sacrificed performance' is broken

For the past two years, the AI industry's "small model" logic was: lightweight = cheap + sacrificed performance. GPT-4o-mini, Claude Haiku, Gemini Flash all followed this thinking—good enough for daily chat, falling short on professional tasks. V4-Flash, with 13B active parameters, ran Pro-level Agent scores, directly stomping this equation into the dirt: 1M context full-repo code analysis, enterprise-grade Agent automation, long document summarization—tasks that previously only Pro could run, Flash can now run them all.

2. The compute math has been recalculated

According to community-circulated comparative estimates (not officially confirmed): V4-Flash running under 1M token context uses only about 0.13 TFLOPs per token and less than 3GB of VRAM; the same input to V3.2 requires about 1.2 TFLOPs and 50GB. If true, Flash saves about 89% compute and about 94% VRAM, allowing roughly 8x more context to be fed for the same money. For enterprise Agent automation companies, cost structure, pricing power, and call frequency ceilings are all reshuffled.

3. Domestic small model strategies forced to reassess

Alibaba Tongyi, ByteDance Doubao, Tencent Hunyuan—everyone's lightweight versions must answer the same question in front of DeepSeek: Your Flash, what makes developers switch over? The seat of "#1 lightweight Agent" has been taken by DeepSeek first.

4. Pulling the rug out from under OpenAI's Codex ecosystem

DeepSeek plugging into Codex using the same Responses API protocol effectively turns OpenAI's monetization entry point into its own customer acquisition channel. Developers can "use OpenAI's posture, call a cheaper Chinese model"—this is more effective than any advertisement.

What does it mean for individual developers?

For individuals, the most concrete sentence from this release is: The scope of 'hard tasks' is shrinking on a monthly basis.

Someone last month just switched their main model from an "expensive one" to a domestic mix, monthly cost dropping from $1200 USD to 4000 RMB. Someone commented then, "Cheap is cheap, but for hard tasks you still need a strong model"—this judgment still holds, but the hardest long-chain tasks are also being caught up on, one by one, by the top-tier players.

Future predictions: What might DeepSeek do next?

The following are all reasonable speculations based on public information, not official promises, for reference only.

1. V4-Pro official version debuts in early August (High confidence) The official statement clearly says "released as soon as possible," and the Responses API documentation also marks Pro support expected to go live in early August. Flash is already like this; the Pro official version will likely refresh the comprehensive capability leaderboard—by then, the tiered logic of "Pro strong, Flash weak" may formally return, or it may be completely restructured.

2. Peak/off-peak pricing officially lands (High confidence) The preview is already hanging on the pricing page; landing is just a matter of time. This also indirectly indicates that DeepSeek's inference load already shows obvious tidal effects, requiring a price lever to shave peaks and fill valleys.

3. Open-sourcing V4 series weights (Medium confidence) DeepSeek has adhered to an open-source tradition since R1 and V3; the community widely speculates the V4 series will eventually be open-sourced. If true, it will reignite a global wave of distillation and fine-tuning, also meaning "Flash-level Agent capability" will rapidly trickle down to any team with a GPU.

4. Tool ecosystem continues to expand (Medium confidence) The Responses API is just the first step. Next up is likely MCP (Model Context Protocol), more IDE plugins, official adaptation for Agent frameworks, turning the "Codex substitute" into an "Agent full suite."

5. Price war enters a new phase of 'not lowering unit price, only lowering task cost' (High confidence) The keyword this time isn't "price cut" but "cache discount + post-training efficiency boost." When a model does the job better at the same price, the unit price doesn't move, but the total task cost drops tangibly—this is a more advanced competitive tactic than direct price cuts, hard for rivals to follow.

Common misconceptions and pitfall avoidance guide

Misconception Correct Understanding Suggestion
"Official version = new model swapped in" Architecture is completely identical to Preview, only post-training was redone, capability improvement concentrated in Agent scenarios Treat as an "upgrade patch," focus regression on Agent workflows
"Score of 82.7 means it can fully replace flagships" All nine benchmarks are Agent/code scenarios; general capabilities like chatting, writing may not be at the same level Choose models by scenario, don't treat a single leaderboard as a full-subject transcript
"Agent Last Exam 25.2 points = model is weak" This is an industry-renowned "hell-level" paper; all models generally score in the 20s-30s, absolute values cannot be compared across benchmarks Look at relative ranking, not absolute score
"Official 9 benchmark scores = ironclad evidence" These are scores self-tested by the official using their own unreleased Harness framework Cross-reference third-party evaluations (AA, Arena), apply a mental discount
"Plugging into Codex costs zero" Protocol is compatible, but the Responses API is a stateless implementation; history must be carried by yourself Check previous_response_id, store parameter limitations
"Cache discount 98% means always cheap" Discount only applies to the hit portion, and peak hours double after peak/off-peak pricing lands Optimize prompt structure to increase cache hit rate, run batch tasks off-peak
"DSBench series scores are authoritative" This is DeepSeek's internal test set, subject to self-testing bias Apply a mental discount when referencing, prioritize international universal benchmarks

Summary

The real signal of DeepSeek V4-Flash official version isn't in those nine benchmark scores, but in the moment these three timelines intersect: old APIs silently offline, Responses API connected by default, Pro not yet officially released.

From today, calling deepseek-v4-flash is no longer "trying it out," but the default option. By September, when various developers make their H2 plans, they'll find their Agent systems deeply bound to this 13B active parameter small model—not chosen, but defaulted.

"If Flash is already like this, what will the Pro official version be like?"—this sentence is probably the shared suspense of the entire industry for the next month.

The only action advice for individuals: Point Codex's interface to DeepSeek and try it for a while. Buy the cheap one to test first; if it's truly enough for daily tasks, then switch your main. If not, you haven't lost much—after all, the most expensive thing is never the model, but your hesitation time.

Reference Links

Note: This article was written on August 1, 2026. Pricing and interface information is subject to the real-time data on the official page. The "Future Predictions" section in the text represents speculation based on public information and does not constitute investment or procurement advice.