TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
DeepSeek V4 Pro Early Morning Surprise: Agent Benchmarks Match Claude Fable 5 Overnight, Entry Fee for Flagship Agents Slashed to a Fraction

DeepSeek V4 Pro Early Morning Surprise: Agent Benchmarks Match Claude Fable 5 Overnight, Entry Fee for Flagship Agents Slashed to a Fraction

2026.08.13
Share
For developers and AI entrepreneurs, the entry fee for flagship agents has been shattered once again.

Author: Claude, TechFlow

TechFlow Editor's Note: In the early hours of August 13, DeepSeek unexpectedly launched the official V4 Pro version: same API name, but Agent capabilities jumped from "basically unusable" in the preview version to rivaling Anthropic's flagship Claude Fable 5, surpassing it on some benchmarks. Even more aggressive is the price, only a fraction of overseas models at the same level. For developers and AI entrepreneurs, the entry fee for flagship agents has been breached again. The only question is, the official price increase notice has already been posted.

In the early hours of today, the DeepSeek official website was quietly updated, upgrading the DeepSeek-V4-Pro model version to the 0813 official version. Developers do not need to change code; calling deepseek-v4-pro accesses the latest model. The new model supports 1 million token context and up to 384,000 token output, with thinking mode enabled by default, and is compatible with OpenAI, Anthropic, and Responses API formats. Mainstream Agent tools such as Codex, Claude Code, and OpenCode can connect directly. The news simultaneously topped the Hacker News homepage (700+ upvotes) and Zhihu trending list (over 20 million heat), becoming the biggest consensus hotspot in today's Chinese and English AI circles.

Early Morning Surprise: Same API Name, Tasks Impossible Yesterday Are Doable Today

The core of this update is Agent capabilities. According to officially published evaluations, the V4 Pro official version's score on its own software engineering benchmark DeepSWE surged from 12.8 in the preview version to 62.7; NL2Repo for natural language to code repository generation rose from 38.5 to 61.5; the full-stack development hard subset DSBench-Hard doubled to 67.2.

To quote a domestic tech media outlet after actual testing: "Same API name, tasks that were basically impossible yesterday can today wrestle with the global first-tier code Agents." In ifanr's actual testing, facing vague instructions like "collect recent tech news to create a homepage mimicking traditional media," V4 Pro could autonomously decompose requirements and execute step-by-step with a high degree of completion. Developers on Hacker News also shared real-world bills: running traffic simulation and distributed physics engine optimization all day, 2 billion tokens cost about $12.5, "found significant performance improvements without introducing any new problems."

Security Offense and Defense Surpasses Fable 5, But the Open Source Throne Isn't Hers Yet

According to official statements, V4 Pro scored 83.3 on the AI security offense and defense benchmark Cybergym, surpassing Claude Fable 5's 83.1; workflow agent AutomationBench also achieved surpassment; high-difficulty reasoning HLE (tools enabled) scored 60.0, second only to Fable 5's 63.0; terminal operation Terminal Bench 2.1 scored 87.9, higher than Claude Opus 4.8's 85.0.

However, two points need to dampen enthusiasm. First, it still has a slight gap with Moonshot AI's Kimi K3 (87.9 vs 88.3 on Terminal Bench 2.1); the throne of the open-source world hasn't changed hands. Second, third-party statements are not as optimistic as the official ones: OpenRouter citing Artificial Analysis gave it a comprehensive intelligence index score of 45.3, "better than 70% of models," showing a discrepancy with the official narrative of "catching up to Fable 5." Official benchmarks were measured under their own Harness framework; independent verification still takes time.

3 Yuan vs 10 Dollars, Flagship Agent's "Fractional Pricing"

What truly makes competitors uncomfortable is the price. The V4 Pro official version maintains preview pricing: 0.025 Yuan per million tokens for cache hit input, 3 Yuan for miss input, and 6 Yuan for output. For comparison, Kimi K3's API price is input $3, output $15; Claude Fable 5 is input $10, output $50. Even considering currency differences, DeepSeek is only a fraction of overseas models at the same level.

Data from OpenRouter shows that due to a cache hit rate as high as 86%, the weighted input price users actually pay is as low as $0.064 per million tokens. To borrow an old saying from the domestic circle: DeepSeek's order is still "those stronger than me aren't as cheap as me, those cheaper than me aren't as strong as me."

Price Increase Notice Posted, API Output "Stiff Phrasing" Criticized

Cheapness may have an expiration date. DeepSeek already announced on August 6 that it would "overall increase API pricing recently, with a large increase," and plans to introduce peak-valley pricing: during peak hours from 9:00 to 12:00 and 14:00 to 18:00 Beijing time daily, V4 Pro miss input will rise from 3 Yuan to 6 Yuan, output from 6 Yuan to 12 Yuan, directly doubling. The official version has not adjusted prices since release, leaving developers a well-understood window period. Another hedging information is that DeepSeek promised during the preview release that Pro prices would be significantly reduced after the mass market launch of Ascend 950 super nodes in the second half of the year. The gap between rise and fall depends on the delivery schedule of domestic computing power.

In addition, developers on X reported significant differences in chain of thought performance between the V4 Pro web end and API calls, with API output phrasing "reading stiffly and having awkward word order," DeepSeek has not yet responded. Teams relying on API for products should test it themselves before launch.

After Flagship Agent Cost Collapse, the Next Highlight is Harness

Putting this matter into a larger picture:

Just this week, research on "stealing reasoning traces from closed-source model APIs" sparked a round of discussion, pointing out cracks in the reasoning moat of closed-source vendors. Now DeepSeek uses "fractional price plus first-tier benchmarks" to prove that the capability gap itself is also being leveled. The remaining pricing power of closed-source flagships has loosened another notch.

For AI application entrepreneurs and investors focusing on the computing power chain, a more worth-worthy foreshadowing is hidden in the official documentation. DeepSeek revealed in the V4 Flash changelog that its Agent operation framework DeepSeek Harness is about to be released, with recruitment and official public account already in place. Industry consensus is that when model reasoning capabilities approach the ceiling, what determines the Agent upper limit will be "office systems" like Harness. The model is already in place; if the framework delivers, the next DeepSeek moment won't be far away.