AI・テクノロジー
DeepSeek V4 Shock | Performance Matches GPT at 1/20th the Price
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・39分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"China's AI has rewritten the world's price list again." — On April 24, 2026, DeepSeek announced V4.
Do you remember the shock of V3 a year ago? That Chinese-made, ultra-affordable AI is back — this time with monster-class specs: 1.6 trillion parameters and a 1-million-token context window.
And it runs entirely on Huawei Ascend chips, with no reliance on American-made NVIDIA hardware.
"Is it really true that it performs on par with GPT-5.5 Pro at 98% less the cost?" We'll unpack that question — covering the latest benchmarks, pricing, and what it means for Japan — in plain, accessible language.
Let's start by getting the announcement straight in five minutes.
On April 24, 2026, DeepSeek published a preview of its next-generation large-scale model, "V4," via its official website and a technical paper. Since the previous model, V3, shook the global AI industry in January 2025, this is "a new generation arriving almost exactly one year later."
Think of it as "a young chef who stunned the world last year confidently unveiling the sequel menu."
In addition to the paper, the model weights (trained files) were uploaded to HuggingFace, making V4 available as an "MIT-licensed open-weight model" — a publicly released AI that anyone can inspect and use commercially.
April 24 became a major announcement that dominated AI industry conversation, with leading financial media outlets including CNBC, Bloomberg, and Fortune all reporting simultaneously.
A defining feature of V4 is that two models are offered simultaneously for different use cases.
"V4-Pro" is an ultra-large model with 1.6 trillion (1,600B) total parameters, activating 49B (49 billion) per inference. "V4-Flash" is a lightweight model with 284B total parameters and 13B active.
Think of it as "V4-Pro being a full-course French dinner, V4-Flash being a fast and satisfying bowl of ramen."
Both use MoE (Mixture of Experts) — an energy-efficient design that only "wakes up" part of the brain when needed. The key insight is that "even though it's a massive 1.6-trillion-parameter model, only 49B parameters actually run at once, making electricity costs and speed realistic."
File sizes are 865GB for V4-Pro and 160GB for V4-Flash. V4-Pro is also referred to as "the world's largest open-weight model" at this point in time.
Both V4-Pro and V4-Flash support contexts of up to 1 million tokens (roughly 600,000–800,000 Japanese characters). Output also supports up to 384K tokens, enabling use cases such as generating lengthy analytical reports in a single pass.
Imagine "swallowing the equivalent of 300–400 A4 pages — two to three paperback books — and then answering questions about them."
A full year of meeting minutes, court documents, hundreds of pages of contracts, entire histories of internal Slack logs, complete codebases — materials that previously had to be fed in chunks can now be submitted in their entirety.
One million tokens matches the industry-leading standard of GPT-5.5 Pro-class models, positioning V4 as a decisive powerhouse for use cases that involve feeding in massive documents for summarization and extraction.
Let's verify not just the specs, but actual intelligence, with numbers.
On the competitive programming site "Codeforces," V4-Pro scores 3206, compared to GPT-5.4's 3168 and Gemini-3.1-Pro's 3052.
"It's a score where a new Chinese entrant in the dojo of the world's top programmers has surpassed the American representative."
On LiveCodeBench (real-world programming tasks), V4-Pro scores 93.5, versus Gemini's 91.7 and Claude's 88.8.
On SWE-Verified (real-world software debugging tasks), V4-Pro scores 80.6, essentially matching Claude at 80.8.
"In the coding domain, the old assumption that Chinese-made means slightly inferior has completely collapsed" — that's a fair conclusion from these results.
On "IMOAnswerBench," a collection of math olympiad-level problems, V4-Pro scores 89.8. With Claude Opus 4.6 at 75.3, Gemini-3.1-Pro at 81.0, and GPT-5.4 at 91.4, V4-Pro ranks second — "far ahead of Claude and Gemini, just barely behind GPT."
"In plain language: an AI capable of reliably solving math at the level needed to pass the University of Tokyo School of Medicine entrance exam has arrived — at an extremely low price." That's the impact.
On the general knowledge test "MMLU-Pro," V4-Pro ties GPT-5.4 at 87.5, trailing Gemini (91.0) and Claude (89.1) slightly. Still, in terms of practical intelligence, the gaps have become extremely small.
On the other hand, on "SimpleQA-Verified," which measures factual accuracy, V4-Pro scores 57.9% versus Gemini's 75.6%.
"In terms of knowledge accuracy, Gemini — backed by Google's vast search data — holds a clear and meaningful lead."
The honest picture of V4 at this stage is "an exceptionally capable model that occasionally states falsehoods with great confidence." The same caveat that applies to other generative AIs holds here: for final decisions in medical, legal, or financial contexts where factual correctness is critical, human review and supplementary external search remain essential.
"Because it's cheap, use the budget saved to build a proper verification process" — that's the practical approach.
V4's real bombshell isn't the specs — it's the price list.
V4-Pro's API pricing is $1.74 per million input tokens (about ¥260) and $3.48 per million output tokens (about ¥520).
For comparison, GPT-5.5 Pro costs $30 for input and $180 for output — meaning "roughly 1/52nd the price for output, and about 1/17th for input."
Compared to Claude Opus 4.7, it's said to be approximately 1/20th the price.
"The equivalent of one Starbucks latte covering an entire month of AI usage costs." That's how dramatic this is.
"Imagine high-quality beef equivalent to premium cuts being sold at a wholesale grocery store for 5% of the price." For developers and companies, this pricing opens up use cases that were previously too expensive to even attempt.
The lightweight V4-Flash costs $0.14 per million input tokens (about ¥21) and $0.28 for output (about ¥42). With a cache-hit design that drops input to $0.028 (about ¥4), "repeating the same prompt becomes essentially free."
"Internal chatbots, automated e-commerce inquiry responses, internal document Q&A, and other high-volume use cases become viable at realistic cost levels."
Calling it "AI that runs on tap water prices" is not an exaggeration. Use cases that had API costs one digit higher in 2025 are expected to turn profitable with V4-Flash for a growing number of businesses.
According to DeepSeek's official paper, V4-Pro achieves just "27% of the floating-point operations (FLOPs) per token and 10% of the KV cache usage" compared to the previous generation V3.2.
"Like a genius student who covers the same material with one-tenth the notes and finishes calculations in one-quarter the time."
This seemingly contradictory goal — making the model larger while reducing operating costs — was achieved through architectural innovation.
"The reason it's cheap isn't that they're selling it at a loss — it's that the operating costs are simply lower." This point, the same as in the DeepSeek V3 era of 2025, continues to draw industry attention.
V4's other major shock is in the chips.
V4 is designed to run inference and training entirely on "Huawei Ascend NPUs." On April 24, Huawei itself announced that "DeepSeek V4 is supported on the latest Ascend AI computing clusters," officially demonstrating that "a frontier-class AI can run at full capacity without NVIDIA's H100 or Blackwell chips."
The fact that "a top-tier AI can be run entirely within China without using the cutting-edge GPUs subject to U.S. export controls" marks what commentators are calling a pivotal moment in the U.S.-China tech competition.
Multiple English-language media outlets report that China's major tech giants — Alibaba, ByteDance, and Tencent — placed large orders for Huawei chips in anticipation of V4's release.
"The move to shift China away from dependence on American NVIDIA and toward a self-sufficient AI economy is taking visible, concrete shape."
Think of it as "moving from importing rice from another country to growing everything in your own domestic paddies." The U.S. side has also reported suspicions that "Blackwell chips may have been smuggled in via third countries and used in training," and the true extent of technological self-sufficiency remains an ongoing point of debate.
DeepSeek and Huawei jointly announced that "V4 delivers equivalent performance on Ascend NPUs and NVIDIA GPUs."
This amounts to an official rebuttal of the industry's long-held assumption that "Chinese chips are several generations behind."
"Like an unknown young pitcher in domestic baseball clocking the same speed as a major league starter." If this claim holds up under independent verification, it would fundamentally call into question the strategic effectiveness of U.S. AI chip export controls — an announcement with implications for global supply chains and geopolitics that could unfold over years.
Let's clarify where V4 wins and where it falls short.
GPT-5.5 Pro is OpenAI's latest flagship, setting the world standard for knowledge depth, conversational quality, and multimodal capability. V4-Pro dominates on price (98% off), wins on coding, is nearly equivalent on math, and is slightly weaker on general knowledge.
"Think of GPT as an all-around private tutor, and V4 as a highly affordable cram school that excels in specific subjects."
For internal development and high-volume API use, choosing V4 makes sense. For everyday conversational use, sticking with GPT is a reasonable split. The larger the company's monthly API bill — ranging from hundreds of thousands to tens of millions of yen — the greater the financial impact of adopting V4.
Claude Opus 4.7 is Anthropic's latest model, well-regarded for coding and agent (autonomous AI worker) capabilities. V4-Pro surpasses it on Codeforces and LiveCodeBench, and costs approximately 1/20th the price — a striking gap.
"If you only need code written, V4 is the clearly rational choice." Claude, however, remains strong in its ability to follow long instructions precisely and in ethical considerations, making it still superior for agentic use cases.
"Use Claude for mission-critical business agents where mistakes are unacceptable, and V4 for bulk generation and prototyping" — that combination may become the norm in the second half of 2026.
Google Gemini 3.1 Pro is unmatched for search integration and multimodal capabilities, scoring 75.6% on SimpleQA-Verified versus V4-Pro's 57.9%.
"As a walking encyclopedia that grasps real-time information from around the world, Gemini is in a league of its own." V4-Pro wins on code generation and reasoning; Gemini wins on factual accuracy of current information.
"Gemini for the latest news, V4 for writing large volumes of code, GPT for general conversation, Claude for business agents" — a four-card lineup that looks set to become the standard in the 2026 AI industry.
What kind of impact will V4 have in Japan?
For Japanese SaaS companies and internal AI deployment teams, V4's price disruption fundamentally changes the calculus around AI spending.
"A calculation where ¥1,000,000 per month in OpenAI fees, maintained at the same usage level with V4, comes to around ¥50,000 per month" is a real scenario.
"The equivalent of moving from a high-end rental to a public housing unit in the same neighborhood, with rent dropping to 1/20th." Companies actively implementing AI in Japan — including Rakuten, LINE Yahoo, Mercari, Sansan, and freee — are likely to accelerate multi-model strategies in the second half of 2026, assigning different models to different use cases.
We have entered an era where comparing V4 as a cost-conscious alternative to GPT/Claude as first choices is simply standard practice.
On the other hand, in high-confidentiality sectors such as Japanese government agencies, finance, and healthcare, caution around direct use of Chinese-made AI remains.
"When using DeepSeek's official service, data may pass through servers in China, making it unsuitable for sensitive use cases from a personal information protection standpoint" — that is the general view. However, since V4 is open-weight (with model weights published under an MIT license), it is technically possible to run V4 independently on your own servers or on AWS or Sakura Cloud.
"Think of it like: the car's blueprints are public, so you assemble it in your own factory and drive it under your own license plate." For enterprise adoption, the choice between official API and self-hosting will be key.
For independent developers and freelancers, V4-Flash is an extremely attractive AI to play with freely.
"With a monthly budget of $5–10, you can build everyday tools — a GitHub Copilot replacement, a translation bot, blog post drafts, data formatting scripts — all in one go." That's the cost level we're talking about.
Japanese language performance has continued to improve since the V3 era, reaching a level approaching ChatGPT.
"Like buying a handy gadget not yet released in Japan from an overseas store — at a price that's still within reach even with the weak yen." Incorporating V4 into side projects, sole proprietorships, blog operations, or study assistance can realistically bring monthly tool costs down to just a few thousand yen.
Sasaki-san runs an apparel e-commerce business with four employees.
"I built an inquiry bot using GPT-4, but ¥120,000 a month was too heavy a burden, so I put it on hold." In May 2026, after switching to V4-Flash, the same quality bot started running for about ¥5,000 per month.
"It felt like gaining a new employee who lives in Slack DMs, handling inventory checks, return policies, and shipping schedules around the clock."
"AI is no longer about mimicking large corporations — it's about choosing something that fits your budget," was the realization.
Nakano-san is in the second year of an information engineering master's program.
"I needed to read 200 English papers before the thesis deadline — it seemed impossible." Using V4-Pro's 1-million-token context window, Nakano-san fed in 10 papers at once as PDFs, extracted key points, and auto-generated comparison tables.
"Meetings with my supervisor transformed into discussion sessions where AI summaries were taken as a given starting point." Monthly API costs came to around ¥800.
"Master's students can now handle the volume of literature that used to be the domain of doctoral students" — "the starting line for research has moved up," Nakano-san says.
Inaba-san works in the DX Promotion Division of a core city.
The seemingly impossible task: "re-read every ordinance revised over the past 20 years and identify all contradictions." Using a self-hosted V4-Pro on AWS, Inaba-san fed in the full text of all ordinances (approximately 3 million characters) in batches, extracting contradiction candidates and checking legal consistency.
"Work that would have taken 10 staff members six months to check through paper ordinance volumes was on track for completion in two weeks."
This became a noteworthy example of a local government leveraging the open-weight advantage — direct use of a Chinese AI service is prohibited, but running the model on the municipality's own servers is permitted.
A. DeepSeek's official web chat (chat.deepseek.com) is free for everyday use. API access for development purposes is paid, but V4-Flash costs just $0.14 per million tokens (about ¥21) — an extremely low price.
"Individuals can get serious, full-featured use out of it for the price of a single coin (around ¥500) per month." For businesses doing large-scale processing, even V4-Pro typically costs less than 1/20th of what competitors charge.
A. When using DeepSeek's official services (web version or API), data may pass through servers in China.
The basic rule is: "Don't put confidential information or personal data into the official service." However, since V4's model weights are published under an MIT license, deploying it yourself on AWS, Azure, or Sakura Cloud allows you to use it without sending data to China.
"You buy the engine, then mount it in your own vehicle" — that's exactly what self-hosting an open-weight model enables, and it's the biggest advantage of this approach.
A. Japanese naturalness has improved significantly since the V3 era. For everyday conversation, blog articles, business emails, and simple code comments, the quality is good enough to feel comfortable.
That said, for fine-grained keigo (honorific) nuances or Japan-specific expressions in specialized fields like law and medicine, V4 still falls a step behind GPT-5.5 Pro and Claude Opus 4.7.
It's similar to "a foreign exchange student who has passed JLPT N1" — "sufficient for everyday business tasks, but a native speaker should polish the final output intended for readers."
A. The basic principle: "precision-critical, complex reasoning, long-document understanding = V4-Pro" and "high-volume, high-speed, cost-sensitive = V4-Flash."
For example, an internal chatbot could handle 99% of questions with Flash, escalating only difficult queries to Pro — a two-tier approach that works well in practice.
"Like commuting by kei car every day and renting a large SUV only for long trips." Many enterprise deployments optimize running costs at a ratio of roughly 90% Flash to 10% Pro.
A. The main options are: "① Use the official DeepSeek API," "② Use via an intermediary service like OpenRouter," or "③ Self-host on AWS, Azure, or Sakura Cloud."
① is the simplest and cheapest, but unsuitable for confidential work.
② costs slightly more but offers flexibility to use multiple models.
③ has higher upfront setup costs but offers peace of mind for information security and compliance.
"For a restaurant, it's like choosing between a franchise contract, going through a distributor, or sourcing directly from the factory." Expect more companies to gradually migrate to option ③ based on their information classification needs.
This is not an announcement that can be dismissed as "cheap Chinese AI showing up again."
A massive 1.6-trillion-parameter model that runs at 1/20th the cost of GPT, outperforms the American top in coding, handles 1-million-token long documents, and operates entirely on Huawei Ascend chips — this is a structural shift in which price, performance, and geopolitics all move at once.
If DeepSeek V3 in January 2025 left the world with the "shock of low-cost AI," then V4 in April 2026 is "the moment Chinese AI reached the frontier and began to shake even America's chip supremacy."
For Japanese companies, the message is that the option to cut AI costs to 1/20th is now a practical reality. For individual users, it is the moment when truly capable AI for around ¥500 a month arrived at a usable level.
Start by using V4-Flash on just one project — that alone is an announcement likely to change the pace of your development six months from now.
This article is a cross-post from AI Friends.
