YAY!!! - DEEPSEEK V4 IS OUT 🚀🚀🚀 Initial benchmarks numbers are ABSOLUTELY ASTOUNDING!! Opus 4.7 Max and GPT 5.5 level! Scrambling to verify their numbers https://t.co/bb7ohcfXZh
A multi-panel infographic: the left panel is a bar chart comparing DeepSeek V4 (Pro-Max and Flash) against Claude Opus, GPT-5.4 and Gemini across coding/reasoning/agent benchmarks (SimpleQA, HLE, Apex, Codeforces, SWE Verified, Terminal Bench, Toolathlon). The right panels plot single-token FLOPs and accumulated KV cache versus token position, visually showing V4’s much lower compute and memory for long contexts — directly illustrating the benchmark & efficiency claims referenced in the tweet (i.e., parity with Opus/GPT-class models and outstanding long-context efficiency).
Source: Hugging Face (deepseek-ai/DeepSeek-V4-Flash model card)
Research Brief
What our analysis found
DeepSeek AI officially released its flagship DeepSeek V4 model series on April 24, 2026, featuring two versions: DeepSeek-V4-Pro (1.6 trillion parameters, 49 billion activated) and DeepSeek-V4-Flash (284 billion parameters, 13 billion activated). Both models support a 1 million token context window and employ a sophisticated Mixture-of-Experts (MoE) architecture alongside innovations including Compressed Sparse Attention, Engram conditional memory, and a Muon Optimizer. The V4-Pro model reportedly requires only 27% of single-token inference FLOPs and 10% of KV cache compared to its predecessor DeepSeek-V3.2 in a 1M-token context setting.
The tweet's claim that DeepSeek V4 reaches "Opus 4.7 Max and GPT 5.5 level" appears to be a slight exaggeration, as most credible comparisons cite Claude Opus 4.5/4.6 and GPT-5.4 as the relevant benchmarks. Nevertheless, leaked internal benchmarks position V4 as a formidable competitor, with claimed scores of over 80% on SWE-bench Verified, up to 98% on HumanEval, and 96% on GSM8K for math and logic tasks. These numbers, if independently verified, would place DeepSeek V4 on par with or ahead of the leading proprietary models from OpenAI and Anthropic.
Perhaps the most striking aspect of this release is the pricing: DeepSeek V4's API is expected to cost approximately $0.28 per million input tokens, making it roughly 50 times cheaper than Claude Opus 4.6's $15 per million input tokens. The model weights are being released under the MIT License on Hugging Face, and the system is optimized to run on domestic Chinese silicon such as Huawei Ascend 950PR chips, as well as consumer hardware like dual RTX 4090s or a single RTX 5090.
Fact Check
Evidence from both sides
Supporting Evidence
Competitive benchmark claims from DeepSeek itself
DeepSeek states its V4 model is "competitive with leading US closed-source models from the likes of OpenAI and Google DeepMind," and the V4-Pro-Max reasoning mode is claimed to be the "best open-source model available today."
Strong SWE-bench Verified performance
Leaked internal benchmarks show DeepSeek V4 achieving over 80% on SWE-bench Verified, which is comparable to Claude Opus 4.5/4.6's verified scores of 80.8%–80.9% and GPT-5.4's approximately 80%.
Exceptional coding benchmark results
DeepSeek V4 claims approximately 90%–98% on HumanEval, compared to GPT-5.4's roughly 92% and Claude Opus 4.5's 88%–92%, supporting claims of frontier-level coding ability.
Math and logic superiority on GSM8K
DeepSeek V4 claims a 96% score on GSM8K, significantly outperforming Claude Opus 4.5's 78.3% on the same benchmark.
Long-context accuracy via Engram memory
The Engram conditional memory system reportedly boosts accuracy from 84.2% to 97% on the Needle-in-a-Haystack benchmark across 1 million tokens, and the model reportedly maintains 100% logical consistency across long contexts.
Dramatic cost advantage over competitors
At approximately $0.28 per million input tokens, DeepSeek V4 is roughly 50 times cheaper than Claude Opus 4.6 and 10 to 50 times cheaper than GPT-5.4, supporting claims of disruptive market positioning.
Contradicting Evidence
Benchmark numbers remain largely unverified by third parties
Multiple sources emphasize that DeepSeek V4's impressive benchmark figures are based on "leaked internal data" and are still "awaiting independent third-party verification," meaning the headline performance claims should be treated with caution.
The tweet uses inaccurate model naming conventions
The claim references "Opus 4.7 Max and GPT 5.5," which do not correspond to standard model names; credible comparisons cite Claude Opus 4.5/4.6 and GPT-5.4, making the tweet's framing a slight exaggeration.
The leap from V3 to V4 is suspiciously large
DeepSeek V3 scored approximately 49% on SWE-bench Verified, and V4's claimed jump to over 80% represents an extraordinary and unusual improvement that experts say warrants significant skepticism until independently confirmed.
Peak performance requires specific configurations
The highest-performing variant, DeepSeek-V4-Pro-Max, is described as a "maximum reasoning effort mode," suggesting that achieving the top benchmark scores may require specific settings and a larger computational budget rather than being the default experience.
No single model dominates across all tasks
Industry analysts note there is no single "best AI model" across all domains; while DeepSeek V4 excels in cost efficiency and context length, Claude Opus 4.6 leads in verified multi-file reasoning benchmarks, and GPT-5.4 offers superior reasoning controls and computer use capabilities.
Report an Issue
Found something wrong with this article? Let us know and we'll look into it.