Opus

Gains

Flash

would

Major

DeepSeek moved its V4-Flash model out of preview and into official public beta on July 31, 2026, and the upgraded version now outscores the company’s own larger, more expensive V4-Pro-Preview model across every agent benchmark DeepSeek published. The catch: it’s the same architecture as before. DeepSeek only retrained it.V4-Flash-0731 keeps the identical 284-billion-parameter mixture-of-experts design and 13-billion active parameters as the preview build — DeepSeek’s changelog is explicit that this is a post-training upgrade, not a new model. On Terminal-Bench 2.1, a measure of complex command-line agent work, it scored 82.7, up from 61.8 for the preview and ahead of V4-Pro-Preview’s 72.1, landing within a few points of Anthropic’s Claude Opus 4.8 at 85.0, according to <a href="https://www.techtimes.com/articles/322513/20260731/deepseek-retrained-v4-flash-beats-its-flagship-pro-nine-agent-benchmarks.htm" rel="noopener noreferrer nofollow">a detailed benchmark breakdown</a>. DeepSWE jumped from 7.3 to 54.4, and Cybergym rose from 38.7 to 76.7.<h2 id="strong-the-comparison-deepseek-wants-you-to-make-and-the-one-it-doesn-t-strong">The comparison DeepSeek wants you to make, and the one it doesn’t</h2>DeepSeek published a nine-benchmark comparison against Opus 4.8. Opus 4.8 leads on all nine, though several gaps are narrow — Agents’ Last Exam sits at 25.2 versus 25.7, essentially a rounding difference. Independent evaluator Artificial Analysis measured the pre-0731 endpoint at 61.8 on Terminal-Bench 2.1 in a July 27 snapshot, matching DeepSeek’s own preview figure to the decimal — solid independent confirmation of the baseline, even if the eye-catching 25.8-point jump some outlets cited compares two different benchmark versions rather than a clean before-and-after.<h2 id="strong-pricing-that-undercuts-everyone-including-openai-s-new-cuts-strong">Pricing that undercuts everyone, including OpenAI’s new cuts</h2>V4-Flash-0731 costs $0.14 per million input tokens on a cache miss (just $0.0028 on a cache hit) and $0.28 per million output tokens — pricing unchanged from the preview, and dramatically below even OpenAI’s newly discounted GPT-5.6 Luna at $0.20/$1.20. The timing is hard to ignore: DeepSeek shipped this release one day after OpenAI cut Luna’s price 80%, positioning V4-Flash-0731 as an immediate counter-move in the same price war rather than a coincidence.<h2 id="strong-what-s-still-unverified-strong">What’s still unverified</h2>As of release, the -0731 weights had not yet appeared on Hugging Face, and DeepSeek’s announced peak/off-peak pricing tiers weren’t live — meaning some details in the official changelog describe a commitment rather than a currently usable fact. Flash also supports 2,500 concurrent requests versus Pro’s 500, a fivefold headroom advantage that likely matters more to production agent fleets than any single benchmark score.The migration cost for existing users is effectively zero: developers already calling the deepseek-v4-flash endpoint get the upgraded checkpoint automatically, under the same model name and API key, with no code changes required. DeepSeek framed this as a deliberate design choice — since the architecture didn’t change, only the post-training pass, there was no technical reason to force a version bump that would fragment the user base across old and new endpoints.<h2 id="strong-what-to-watch-next-strong">What to watch next</h2><ul><li style="font-weight: 400;">Whether the -0731 weights land on Hugging Face as promised, enabling self-hosted deployment.</li><li style="font-weight: 400;">Whether DeepSeek’s separate V4-Pro official release, still pending, closes more of the gap with Opus 4.8.</li><li style="font-weight: 400;">Whether rival labs respond with their own retrained, price-unchanged upgrades rather than net-new model launches.</li></ul><h2 id="strong-sources-strong">Sources</h2><ul><li><a href="https://www.techtimes.com/articles/322513/20260731/deepseek-retrained-v4-flash-beats-its-flagship-pro-nine-agent-benchmarks.htm" rel="noopener noreferrer nofollow">DeepSeek Retrained V4-Flash Beats Its Flagship Pro on Nine Agent Benchmarks — Tech Times</a></li><li><a href="https://benchlm.ai/blog/posts/deepseek-v4-flash-0731" rel="noopener noreferrer nofollow">DeepSeek V4 Flash 0731: Benchmarks, MIT Weights, and Pricing — BenchLM.ai</a></li><li><a href="https://www.marktechpost.com/2026/07/31/deepseek-upgrades-deepseek-v4-flash-0731-with-major-agentic-and-coding-gains/" rel="noopener noreferrer nofollow">DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains — MarkTechPost</a></li></ul>Disclaimer: This content is meant to inform and should not be considered financial advice. The views expressed in this article may include the author’s personal opinions and do not represent Times Tabloid’s opinion. Readers are advised to conduct thorough research before making any investment decisions. Any action taken by the reader is strictly at their own risk. Times Tabloid is not responsible for any financial losses.The post <a href="https://timestabloid.com/deepseeks-cheap-model-just-beat-its-own-flagship-on-nine/" rel="noopener noreferrer nofollow">DeepSeek’s Cheap Model Just Beat Its Own Flagship on Nine Benchmarks</a> appeared first on <a href="https://timestabloid.com" rel="noopener noreferrer nofollow">Times Tabloid</a>.

DeepSeek’s Cheap Model Just Beat Its Own Flagship on Nine Benchmarks

Table of Contents

Times Tabloid

Other articles published on Aug 2, 2026

Silver Price Warning: Don’t Fall for $500 Hype Today – Here’s the Real Chart

Xaman Wallet Issue Warning as Fake Token Scams Threaten XRP Community

U.S. ETF Launches Hit Record Pace as Leveraged Products Surge

DEX Spot Volume Hits Record 24% of CEX Trading, Highest Since 2019

Rising JGB Yields Put Japan’s Debt Strategy and Global Markets at Risk

Astarter Joins BitBall to Accelerate AI-Driven Sports Intelligence on BNB Chain

Join our free newsletter for daily crypto updates!