DeepSeek V4 Rumors Escalate: Challenging Frontier Models with Disruptive Pricing

Between July 18 and 20, rumors surrounding DeepSeek V4 significantly escalated. Multiple leaks suggest the model is in gray-box testing and slated for an imminent release. It has garnered immense attention because the circulating narrative portrays V4 as a direct competitor with a core value proposition of "high performance at an extremely low price," potentially disrupting the closed-source API pricing paradigm and forcing companies like OpenAI to rethink their rate limits and high-margin cycles.

Key Rumors and Details

Sources like Xin Zhiyuan indicate that the official release will feature two tiers: V4 Flash and V4 Pro. The chain-of-thought opener has reportedly changed from "Let me" to "I'm/I'll," with the introduction of peak/off-peak pricing. Performance benchmarks vary: some leaks suggest it reaches Claude Opus 4.8 levels, while others cite a SWE-bench score of 80.6% (close to Opus 4.6's 80.8%), with an output price of $3.48/M—about one-seventh of the latter—and a 1M context window. Shared test screenshots highlight major improvements in coding, agentic tasks, and 3D generation, such as generating an FPV flight simulator with SO(3) pose control in one pass, drawing comparisons to GPT-5. A repost by @TinfoilTricorn mentioned potential new KV management technology, preliminarily judged to be superior to Kimi K3.

Community Reactions and Debates

@teortaxesTex noted that DeepSeek is expanding like a true AGI company, predicting V4 will be the strongest Chinese LLM, likely reaching a semi-frontier level. However, based on current tests, he believes V4's absolute scale might be slightly inferior to Kimi K3, though still a "revolutionary" step due to its RL compute advantages. @bindureddy argued that DeepSeek might deliver a model close to Kimi K3's quality at a tenth of the price, eroding closed-source moats. @Linahuaa expects strong open-source models to end the era of overpriced closed-source APIs.

Controversies and Uncertainties

All materials currently stem from unverified leaks and commentary, without official confirmation from DeepSeek. @teortaxesTex clarified that the recent "gray-box testing" under the suspected codename "Fable" is not an official release; the community suspects some vendors might be secretly collecting data or testing unreleased models via API with safety guardrails, which could damage user trust if mishandled. Furthermore, the release timeline is based solely on a "next Monday" rumor. Regarding its capabilities, the narrative remains unresolved: some claims assert it directly rivals Kimi K3 and Opus 4.x, while others suggest it falls "slightly below Kimi K3."

2026-07-18 ~ 2026-07-20 · 15 related posts