RUMOR: one post named the model half the Chinese AI market has been waiting for

Editorial label: RUMOR. On September 7, 2026, leak tracking account Mr_Salio posted on X that DeepSeek is preparing V5, described as a full rebuild of the model's foundation rather than an incremental update to V4. The core claim is twofold: V5 would target Mythos class capability, Anthropic's most restricted and advanced tier, and parity with OpenAI's GPT-6 Astra, delivering that at a much lower cost. The post was reposted on Mastodon hours later and picked up by aggregators including CometAPI and kie.ai. DeepSeek has confirmed nothing, and there is no model card, published benchmark, or API identifier for the supposed V5.

Why the specific 'Mythos class' claim carries more weight than ordinary hype

Calling your own model a Mythos rival is no small claim: Anthropic described Mythos as a noticeable improvement over its own flagship on internal coding and research benchmarks, and keeps access restricted to approved programs. If DeepSeek is genuinely aiming at that tier, V5 would mark the first time a Chinese lab publicly stated, even through an unofficial leak, intent to match Anthropic's highest and least accessible level, not just the commercial top of Western competition.

DeepSeek's release pattern doesn't favor pre-announcement

It's worth weighing who is making the claim. DeepSeek has a track record of quiet launches, often timed for weekends, without prior marketing campaigns. The company's own official blog, checked during this reporting round, mentions V5 nowhere: the most recent entry is DeepSeek-V4-Flash-Vision-Exp, published September 1 as the first experimental multimodal model in the V4 family, built on V4-Flash-0731. That doesn't invalidate the rumor, but it confirms that, so far, every public signal comes from outside the company, never from inside it.

The context that lends partial credibility: the V3 to V4 gap

There is a precedent supporting the from scratch rebuild hypothesis. The jump from V3 to V4 also took longer than DeepSeek's usual cadence, and part of the delay was attributed to re engineering the model to run natively on Huawei silicon, a process the technical community described as a ground up rewrite. If V5 is repeating that kind of structural rework, a longer gap and the absence of any pre announcement so far would be consistent with the company's past behavior, even without proving the date or the Mythos comparison.

The missing piece: no numbers, no weights, no reproducible proof

Nothing circulating today is independently verifiable. There is no parameter count, no benchmark reproduced by a third party, no weights published on Hugging Face, and no API price. A minority reading, also unconfirmed, describes A/B tests focused on visual coding workflows compared against Anthropic's Fable 5, which suggests either a different variant of the main rumor or simple noise from multiple accounts trying to predict the same launch with different details.

What changes in MaxAssistant's read

This rumor is worth logging because it has a traceable origin and because DeepSeek has already proven capable of generational leaps that catch the market off guard. But the claim of beating Mythos is, today, a leak account's bet amplified by aggregators, not technical evidence. Anyone tracking the open weight model market should keep V5 on the radar as a genuine medium term possibility, without treating 'Mythos level' as a specification, and wait for the first verifiable signal: an API identifier, a Hugging Face entry, or a direct announcement from DeepSeek itself.

Sources

Salio (@Mr_Salio) on X: https://x.com/Mr_Salio/status/2093668379118719403 | CometAPI, DeepSeek V5: Expected Specs, Features and Performance vs GPT-6: https://www.cometapi.com/deepseek-v5/ | kie.ai, What Is DeepSeek V5? Personal-Agent Focus: https://kie.ai/blog/what-is-deepseek-v5