DeepSeek-V4-Pro Production Release – Performance Closing in on Fable 5 at 1/60th the Price

Overview#

Late on August 12, 2026, DeepSeek quietly upgraded the V4 Pro model from preview to production. The update arrived without a formal launch event or release notes – the only indication was an update to the model version identifier in the API pricing page, now showing DeepSeek-V4-Pro-0813. The model name deepseek-v4-pro remains unchanged, so existing users can start using the new version without any code changes.

Earlier that same day, Elon Musk’s xAI also officially released Grok 4.6, which scored 1753 Elo on the GDPVal-AA v2 leaderboard, surpassing Fable 5. The simultaneous release of two high‑value flagship models drew widespread industry attention.

This article analyses the V4 Pro production upgrade across several dimensions: performance gains, benchmark results, cost‑effectiveness, and selection guidance between Flash and Pro. It should be noted that as of the time of writing (August 13, 2026), DeepSeek has not yet published a complete changelog for V4-Pro-0813. Benchmark data in this article comes primarily from community sources and unofficial channels – official figures should be awaited for confirmation.

V4 Series Parameter Recap#

To better understand where this upgrade fits, here’s a quick recap of the V4 series specifications:

Key takeaway: The model architecture has not changed. 1.6T total parameters, 49B active parameters, MoE architecture, 1M context window, and 384K max output length – all these hardware‑level specs have remained unchanged since the April 24 preview release. Similar to the Flash production release on July 31, the Pro production performance gains come primarily from extensive post‑training tuning.

In other words: the Preview version was the April baseline, Flash production was the July‑tuned lightweight version, and this Pro production release is the August‑tuned flagship version. Flash production had previously surpassed Pro Preview on Agent benchmarks; with Pro now receiving the same post‑training treatment, the performance gap has been corrected.

Ten Benchmark Comparisons: Preview vs Production#

The following table shows cross‑verified benchmark comparisons. Note that Preview numbers are official scores from the April release, while Production numbers are from leaked community sources for the 0813 version – no independent third‑party reproduction results are available yet.

Key highlights:

• DeepSWE jumped nearly 5× (12.8 → 62.7). This is DeepSeek’s in‑house software engineering benchmark, evaluating the model’s ability to modify code and fix bugs in real‑world engineering scenarios. The Preview’s 12.8 was low; the Production’s 62.7 surpasses Anthropic’s previous‑generation flagship Opus 4.8 at 58.0 – a qualitative leap.
• Terminal Bench 2.1 hit 87.9. This widely cited Agent benchmark evaluates the model’s ability to execute terminal commands, modify configurations, and fix issues. The Production score of 87.9 is just 0.1 points behind the top‑ranked Fable 5 at 88.0.
• Cybergym surpassed Fable 5: 83.3 vs 83.1. This benchmark tests cybersecurity attack/defence capabilities – surpassing the closed‑source leader in this domain is a significant milestone.
• Automation Bench and DSBench series doubled: automation workflows and full‑stack development capabilities both saw multiplicative gains – a generational leap.

Overall, the performance gap between Preview and Production is substantial.

Horizontal Comparison: Against Flash Production and Competitors#

The following table shows leaked cross‑benchmark comparisons for the 0813 release against mainstream flagship models including Fable 5, Opus 4.8, GLM‑5.2, and Kimi K3:

The data shows a clear profile for Pro Production – it may not top every single benchmark, but it sits firmly in the top tier across all of them. It trails Kimi K3 by just 0.4 points on Terminal Bench (88.3 vs 87.9) but leads Opus‑4.8 (85.0) by 3 points; it takes first place on Cybergym at 83.3; both DeepSWE and NL2Repo are among the top; DSBench‑Hard at 67.2 is second only to Opus‑4.8’s 71.7.

“No obvious weak spots” is a core competitive advantage for Agent models. Agent tasks aren’t single‑turn Q&A – they involve multi‑step long‑chain execution: terminal operations, code modification, tool calls, data parsing, environment setup, and more. A deficiency in any single capability can cause the entire chain to fail. The Preview’s weak performance on long‑code tasks like DeepSWE (12.8, indicating a high failure rate) was a significant limitation; with that gap now addressed, the model’s real‑world Agent usability has improved dramatically.

It’s also worth noting that the performance gap between Flash Production and Pro Production has returned to “Pro slightly ahead” – about 5–10 points across benchmarks. Despite having a fraction of Pro’s parameters (284B vs 1.6T total, 13B vs 49B active), Flash remains exceptional value.

The Technical Path Behind the Performance Gains#

Consistent with the Flash production upgrade, Pro’s performance gains come primarily from post‑training optimisation.

DeepSeek hasn’t published a technical report for Pro Production yet, but based on Flash’s changelog, it’s reasonable to assume the same upgrade strategy was applied – the model architecture, parameter count, and active parameter count were all left untouched, while the post‑training pipeline was thoroughly reworked. To use an analogy: it’s the same engine, but the transmission, ECU calibration, and suspension have been retuned – delivering better performance without hardware changes.

This suggests that under the MoE + post‑training technical roadmap, there’s still significant room to push the performance ceiling. Especially for long‑chain Agent scenarios, the model architecture itself is already capable – the bottleneck lies more in large‑scale post‑training tailored to code toolchains, terminal operations, and multi‑step reasoning. Training data quality and tuning strategies may have a greater impact on final performance than simply adding more parameters.

For context, here’s a quick recap of V4’s core architecture:

• MoE (Mixture of Experts): The 1.6T parameters are split into thousands of “expert sub‑models”; each forward pass activates only 64 experts (totalling 49B parameters) – meaning the model has the knowledge capacity of 1.6T but the inference compute cost of a 49B model.
• Hybrid attention (CSA + DSA + mHC): Reduces computational overhead in long‑context scenarios – at 1M tokens, per‑token inference cost drops to about 10% of V3.2, and KV Cache usage is significantly lower.
• Engram memory technology: Improves long‑context precision recall – better at retrieving details from large codebases and long documents.

New Capabilities: Multimodal Support, Responses API, and Codex Integration#

The Preview was text‑only; Production natively supports image reasoning for the first time. According to 36Kr’s hands‑on testing, the DeepThink engine can analyse screenshots, interpret images, and process mixed text‑image documents within a single reasoning flow. For example, submitting an error screenshot allows the model to directly read the code, error messages, and stack trace from the image and propose a fix – eliminating the need for separate OCR preprocessing in development workflows.

On the API side, Pro Production shares the same interface capabilities as Flash Production:

• Native Responses API (OpenAI‑compatible) support – Codex tools (Codex CLI, ChatGPT desktop, VS Code Codex plugin) can be integrated by simply changing the base_url, with the same automated setup script as Flash
• Also supports Anthropic API format – Agent tools like Claude Code and OpenCode can switch backends by changing just one line of code
• Thinking mode toggle (enabled by default), JSON Output, Tool Calls, conversational prefix completion (Beta), and FIM completion (Beta, non‑thinking mode only) are all supported

In short, if you already have code calling Flash, migrating to Pro is as simple as changing the model parameter from deepseek-v4-flash to deepseek-v4-pro.

Pricing & Concurrency: 6 CNY/MT Output, Three Times Flash#

The following pricing is from the official pricing page as of August 13, 2026:

Pro is priced at about three times Flash – input cache miss rises from 1 CNY to 3 CNY, output from 2 CNY to 6 CNY, with cache‑hit pricing much closer. The concurrency difference is even more pronounced: Flash at 2500 vs Pro at 500 – consistent with their positioning (Flash for high‑volume requests, Pro for high‑value long tasks).

Peak‑hour pricing doubles (the time‑of‑use mechanism applies to Pro as well): 9:00‑12:00 and 14:00‑18:00 Beijing time (7 hours per weekday) – all pricing items double, meaning Pro output during peak hours is 12 CNY per million tokens.

International competitor pricing comparison (converted to USD):

V4 Pro’s output price is about 1/28 of Opus 4.8 and 1/57 of Fable 5 – roughly on par with Grok 4.6’s output price ($6). Combined with its Terminal Bench 2.1 score just 0.1 points behind Fable 5, the claim of “performance closing in on Fable 5 at one‑sixtieth the price” is well grounded.

Two additional points:

1. DeepSeek has already posted a price increase notice on the pricing page: “We plan to raise DeepSeek API pricing in the near future, with a significant increase expected – please plan your usage accordingly.” The current pricing window may be limited.
2. The SWE‑bench Verified 80.6% figure has not been included in the main analysis, as it has only appeared in encyclopaedia entries and third‑party articles without official confirmation – treat it as unofficial until formally published.

Flash vs Pro – Selection Guide#

Bottom line: For 95% of daily tasks, Flash Production is more than sufficient. Pro’s 3× price doesn’t translate to a 3× experience improvement in most scenarios. Pro’s strengths are concentrated in long‑chain Agent tasks, high‑difficulty coding, and complex data science – and for many users, the “small difference” they perceive is simply because their tasks aren’t demanding enough to reveal the 5‑point gap.

Additionally, according to 36Kr’s hands‑on testing, there are two practical considerations:

1. In large‑scale code generation, Thinking mode can consume 80%+ of output tokens on reasoning, compressing the actual content. In BFS/DFS/A* visualisation tests, Thinking mode truncated output to about half, while disabling it allowed full execution in 54 seconds.
2. Third‑party intelligence scores show divergence: Artificial Analysis gives V4‑Pro‑0813 an intelligence index of 53, only 1 point above Flash‑0731’s 52 – reinforcing the view that “the gap isn’t noticeable in everyday scenarios.”

Also worth watching is the DeepSeek Harness inference framework, currently in closed beta. The Harness team opened beta signups on August 1, with early testers reporting strong results. The framework’s impact on model performance may be larger than expected – whether the benchmark scores above are boosted by Harness optimisation is currently unconfirmed. Once Harness is publicly released, Pro’s practical performance may have further room to improve.

Concurrency, API Migration, and Practical Tips#

1
from openai import OpenAI
2

3
client = OpenAI(api_key="<your_api_key>", base_url="https://api.deepseek.com")
4

5
# Original Flash call
6
response = client.chat.completions.create(
7
    model="deepseek-v4-flash",
8
    messages=[{"role":"user","content":"Write a quicksort implementation"}],
9
    temperature=1.0,
10
)
11

12
# Switch to Pro – all other parameters stay the same
13
response = client.chat.completions.create(
14
    model="deepseek-v4-pro",
15
    messages=[{"role":"user","content":"Write a quicksort implementation"}],
16
    temperature=1.0,
17
)

1
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex

Pricing Trends and Industry Context#

The “significant API price increase” notice on the pricing page is worth paying attention to. DeepSeek posted a related announcement on August 6, and the Pro production release didn’t coincide with a price adjustment – the length of this low‑price window is uncertain. DeepSeek has recently raised substantial funding at a high valuation, but compute costs remain a core constraint – V4 Flash alone has surpassed 4.66 trillion tokens in weekly calls on OpenRouter, topping the global charts for several consecutive weeks. The pressure behind such massive usage at consistently low prices is real.

Time‑of‑use pricing is essentially resource scheduling via price signals: when compute is scarce during weekday daytime hours, prices rise, shifting non‑urgent batch workloads to off‑peak periods to ensure stability for finance, development, and real‑time Agent scenarios. For developers, rather than worrying about price fluctuations, it’s better to build tiered routing and caching optimisation into your architecture – even if prices rise later, well‑prepared users can still keep costs under control.

Summary#

V4 Pro Production represents a substantive generational leap, not a minor iteration: DeepSWE improved nearly 5×, Terminal Bench hit 87.9 (just 0.1 points behind Fable 5), Cybergym surpassed Fable 5, and multimodal and Codex support have been added. Combined with the 6 CNY per million tokens output price, it’s one of the most cost‑effective flagship‑class Agent models available today – far cheaper than comparable overseas models.

But Pro isn’t the best choice for every scenario: For everyday tasks, Flash Production already outperforms Pro Preview, at one‑third the cost and five times the concurrency. If Flash already meets your needs, don’t pay extra just for benchmark numbers. Only when you encounter long‑chain Agent tasks where Flash frequently fails should you consider switching to Pro.

A final note: this article was written on August 13, 2026. Some data comes from community sources and third‑party reports – DeepSeek has not yet published a complete changelog. Treat benchmark numbers as references, and validate performance against your own use cases.

With Pro Production now live, the V4 series product matrix is largely complete – Flash offers extreme cost‑effectiveness, Pro delivers flagship performance, with output pricing ranging from 2 to 6 CNY per million tokens, further lowering the global AI API pricing bar. For developers, this trend is broadly positive.

References#

Information compiled from the following public sources. These are from the internet and not guaranteed to be fully accurate – for reference only. Prices, policies, and benchmark methodologies are subject to change – always refer to DeepSeek’s official API documentation and latest announcements.

• DeepSeek Official Website (Homepage announces V4 Pro production release)
• DeepSeek Official API Pricing Page (Models & Pricing)
• Breaking: DeepSeek V4 Pro Officially Released (51CTO / Su San)
• DeepSeek V4 Pro Production: 1M Context + 384K Output, 3× More Expensive Than Flash
• Hands‑on: Production DeepSeek V4 Pro – Agent and Vision Capabilities (GeekPark @ 36Kr)
• Liang Wenfeng vs Elon Musk – DeepSeek V4 Pro Puts Pressure on the Global Leader (Phoenix Tech @ 36Kr)
• DeepSeek V4 Pro Production Hands‑on: Agent Capabilities Up 5×, Zero‑Cost API Migration Guide (CSDN Beijing Dev Community)
• DeepSeek V4 Pro Released: Everything I Could Want in a Model, Plus Multimodal (Jingxuan AI @ 36Kr)
• DeepSeek V4 Pro Goes Production – Performance Nearing Fable 5 at One‑Sixtieth the Price (Alibaba Cloud Dev Community)
• DeepSeek V4-Pro-0813 Released: Performance Near Fable 5, But Mixed Real‑World Results – Value Questionable (ZOL AI)
• DeepSeek V4 Production Release Imminent – API Peak‑Hour Prices to Double (Interface News @ Sina Finance)
• DeepSeek-V4-Pro Encyclopedia Entry (Timeline and Pricing)