DeepSeek V4 Launches: 1.6 Trillion Parameters, 8x Cheaper Than GPT-5.5

Today (April 24), DeepSeek released V4.

Bloomberg's headline read "Upending Silicon Valley Again a Year Later." Whether that's an exaggeration aside, this launch actually bundles two models:

  • V4-Pro: 1.6 trillion total parameters, 49B activated, 1M token context window
  • V4-Flash: 284 billion total parameters, 13B activated, also 1M token context

Both are open-sourced under the MIT license, with weights uploaded directly to HuggingFace.

Benchmarks First

On the competitive programming Codeforces rating, V4-Pro scored 3206. GPT-5.4 scored 3168 — the closed-source flagship that was just replaced by GPT-5.5 yesterday.

An open-source model has caught up in competitive programming.

Other numbers:

Benchmark V4-Pro Score
LiveCodeBench 93.5%
HMMT 2026 Math Competition 95.2%
IMOAnswerBench 89.8%
MMLU-Pro 87.5%

But it didn't win everything: SWE-bench Pro (real-world software engineering tasks) came in at 55.4%, while Kimi K2.6 scored 58.6%. It won competitive programming but lagged by 3 points in engineering tasks.

DeepSeek's official statement: "DeepSeek-V4-Pro-Max achieves a significant breakthrough in knowledge capabilities among open-source models, firmly establishing its position as the strongest open-source model today."

Pricing: An Order of Magnitude

This is what really gets attention.

Version Standard Input Cache Hit Input Output
V4-Flash $0.14/M $0.028/M $0.28/M
V4-Pro $1.74/M $0.145/M $3.48/M

Yesterday GPT-5.5 launched with standard output at $30/M and Pro at $180/M. Claude Opus 4.7 output is $75/M.

V4-Pro output is $3.48/M.

That's 8.6x cheaper than GPT-5.5 standard and 21x cheaper than Claude Opus 4.7.

This isn't a discount — it's an order-of-magnitude gap. For high-volume teams, this difference can determine whether a product is profitable.

How It Gets So Cheap: Architecture

V4 introduces a hybrid attention mechanism (CSA + HCA): Compressed Sparse Attention handles global dependencies, Heavily Compressed Attention handles local information, used in combination.

In the 1M-token long-context scenario, V4-Pro requires only 27% of the inference compute and 10% of the KV cache compared to V3.2.

Three additional improvements were applied:

  • FP4 + FP8 mixed-precision training, with expert layers using FP4 and other parameters using FP8
  • Manifold-Constrained hyperconnections to strengthen residual signal propagation
  • Muon optimizer to improve convergence speed and training stability

Pre-training data exceeds 32T tokens.

This efficiency isn't from parameter tuning — it's architectural compression of computation, which then flows through to pricing.

The Timing Is Interesting

GPT-5.5 launched yesterday; DeepSeek V4 launched today.

These two events together force enterprise customers to recalculate. The performance gap still exists, but the price gap has grown large enough to demand serious attention. Especially:

  • For high-volume teams: $3.48 vs $30 output cost is a number that can decide architecture selection
  • For those needing open source: MIT license, public weights, no barriers to private deployment
  • For Chinese enterprise users: DeepSeek has fewer concerns around compliance and data sovereignty

Last month there were reports that Tencent and Alibaba were in talks to invest in DeepSeek, with a valuation exceeding $20 billion. A company that refused VC for three years is now talking about money — V4's launch is likely not just a product milestone but a proof-of-capability move ahead of fundraising.

Whether it can convert this price advantage into scale will be answered in six months.

Sources: DeepSeek-V4-Pro Model Card (HuggingFace, deepseek-ai, April 24, 2026); DeepSeek V4 Released: Open-Source 1.6T MoE, 1M Context (ofox.ai, April 24, 2026); CocoLoop; DeepSeek Unveils Newest Flagship AI Model a Year after Upending Silicon Valley (Bloomberg, April 24, 2026)