DeepSeek V4 Full-Power Version Expected Tomorrow – What to Expect

The MLNLP community, a well‑known forum for machine‑learning and natural‑language‑processing researchers and practitioners, announced that DeepSeek V4 (GA) will be released as early as tomorrow, with the latest rollout no later than a few days.

Two versions are available: DeepSeek V4 Flash and DeepSeek V4 Pro . Early gray‑scale testers have already received access.

To identify whether you have been gray‑scaled, the community suggests checking the first‑person token in the model’s chain‑of‑thought output: if it starts with "I'm" or "I'll" instead of the older "Let me", you are likely using the V4 GA version.

Performance impressions from early adopters

Overall performance approaches Opus level 4.8 and the encoding quality rivals GPT‑5.6 Sol.

Agent capabilities see a large boost, with noticeably better 3D and SVG generation.

For the same task, V4 requires more iteration steps than Fable 5.

One developer, Pankaj Kumar, noted that while V4 may not beat the newly released Kimi K3, its price is expected to be significantly lower, potentially creating another "DeepSeek moment".

First round of demos

External demos of the GA version have begun leaking, eliciting mixed reactions: some claim V4 can match Claude 5, while others observe little difference between the Pro and Flash editions.

Examples include a 3D‑style shooting game generated by V4 Pro, a hybrid "Minecraft + No Man's Sky" HTML game, and a classic "Cut the Rope" clone, all produced without additional coding.

Pricing and the new "peak‑valley" billing

DeepSeek introduced a tiered pricing scheme after the GA launch:

deepseek‑v4‑pro : $0.87 per million output tokens during off‑peak, $1.74 during peak; uncached input costs $0.435 per million tokens.

deepseek‑v4‑flash : $0.28 per million output tokens off‑peak, $0.56 during peak; cached input costs $0.0028 per million tokens.

While the change has little impact on individual users, teams that run agents continuously will need to recalculate costs. Cached‑hit pricing remains extremely low, allowing batch tasks and data generation to stay affordable even during peak periods.

Compared with Fable 5’s $50 per million output tokens, V4 still offers the best cost‑performance ratio.

Benchmark results

In DeepSeek’s internal SWE‑bench Verified test, the V4‑Pro‑Max model trails Claude Opus 4.6 Max by only 0.2 percentage points, yet its price is roughly one‑seventh of Claude’s.

However, Opus retains an edge in long‑context retrieval, specialized software‑engineering tasks, and certain knowledge‑reasoning benchmarks.

Model deprecation

Effective July 24, the older models deepseek‑chat and deepseek‑reasoner will be retired.

Overall, the community’s three‑month wait for the "full‑power" DeepSeek V4 appears to be ending, but the model is unlikely to become the absolute top performer across all metrics; its main advantage lies in delivering Opus‑level capabilities at a fraction of the price.