Rumor tracker · 2026-08-22

DeepSeek V5 Status Tracker
September rumors, official silence

After V4 Pro's official release landed 08-13 and the API price hike took effect 08-17, rumors of 'V5 in September, Mythos-class' started circulating. DeepSeek has said nothing. This page tracks only verifiable facts.

#deepseek v5#september rumor#V4 price hike#V4 Pro on sale

Current status at a glance

Rumor

V5's current phase

In mid-August the web circulated 'September release, Mythos-class performance.' Zero official word from DeepSeek.

08-13

V4 Pro official GA

1.6T-class MoE, 1M context, 384K max output, with native Responses API and Codex support.

08-17

API price hike effective

Peak/off-peak pricing landed: V4-Pro peak output ¥27 per million tokens (+350%), cached input up as much as 1100%.

On sale

The V4 series on QCode

deepseek-v4-pro and deepseek-v4-flash are both callable; catalog prices are live.

Where the V5 rumor comes from

The source is industry interpretation after the 08-17 price hike took effect: in the price-change announcement, DeepSeek mentioned it 'will develop stronger large models,' and media stitched that together with supply-chain whispers into 'V5 in September, benchmarked against Mythos 5.' Note: the announcement itself never says 'V5' — 'September' and 'Mythos-class' are both secondhand extrapolation.

The V4 line is still iterating fast

While waiting for V5, the V4 line hasn't stopped: on 08-13 V4 Pro went GA (stronger agentic capability, 87.9 on Terminal-Bench 2.1); on 08-21 the multimodal V4-Flash-Vision-Exp went live (images billed as tokens, same price as V4-Flash). DeepSeek's rhythm is to lay the V4 series out fully first — the next generation comes after.

Timeline

2026-08-06

The developer console warns of 'a broad, significant API price increase soon.'

2026-08-13

V4 Pro official GA + the price-change plan published: peak/off-peak pricing, peaks 9:00-12:00 / 14:00-18:00 (Beijing time).

2026-08-17

New prices take effect; the 'V5 in September' rumor starts circulating (born from reading into the 'develop stronger models' phrasing).

Confirmed vs rumors

Confirmed

V4 Pro's official GA on 08-13, the peak/off-peak price hike effective 08-17, and V4-Flash-Vision-Exp going live 08-21 — all have official announcements or Caixin-level reporting behind them.

Rumors

'V5 in September' and 'Mythos-class' have no official source — they're amplified readings of the phrase 'develop stronger large models' in the price-hike announcement. Even the name V5 itself is unconfirmed.

Wait for V5 or use V4

Wait for V5

Fits: non-urgent needs. Risks: no date, no specs — and a new generation will most likely inherit the new pricing regime.

Use V4 Pro / Flash

In stock, open weights, a proven 87.9 on TB2.1. Off-peak prices are still half of peak — scheduling alone keeps costs down.

How to follow along

DeepSeek only recognizes two channels: the official docs site (api-docs.deepseek.com) and the official WeChat account. If V5 exists, it will show up in one of those two places first. This page tracks in sync.

On QCode

There is no callable version of V5. QCode sells deepseek-v4-pro (catalog price $0.66/$1.98 per million tokens) and deepseek-v4-flash ($0.22/$0.66); the official peak/off-peak price gaps flow through to the catalog, and the console shows live terms.

FAQ

Has DeepSeek V5 been released?

No. As of 2026-08-22 there is zero official V5 news — 'September release' is media speculation.

What's the basis for the V5 rumor?

A single official phrase: the price-hike announcement mentioned 'will develop stronger large models.' Everything beyond that is speculation.

How much did V4 prices rise?

V4-Pro peak output is ¥27 per million tokens (about +350% over before); cached input rose as much as 1100%; V4-Flash rose up to 400%. Off-peak is half of peak.

Is DeepSeek still cheap after the hike?

Still the cheapest tier among open-source flagships: off-peak output is about half of peak, and stacked with the cache-hit discount, real costs stay far below closed-source flagships.

How does V4 Pro GA differ from Preview?

The GA build (0813) significantly strengthens agentic capability, natively supports the Responses API and Codex, and scores 87.9 on TB2.1 by official terms.

What is the multimodal V4-Flash-Vision?

An experimental build launched 08-21: images are billed as tokens (up to 384 tokens per image), priced identically to the regular V4-Flash.

Sources

DeepSeek's official price-change announcement (2026-08-13, via Xinhua Finance), guancha.cn (08-17), Caixin (08-21), Kaiyuan Securities commentary (08-06), Mydrivers (08-18, the V5 rumor).

Until V5 arrives, V4 is in stock

deepseek-v4-pro / flash are on sale at QCode — off-peak costs half of peak.