AI・テクノロジー
DeepSeek "R3" Officially Released — Surpasses o3 in Math and Coding Reasoning at 1/8 the Cost
機械翻訳 / Machine-translated
On August 29, 2026, DeepSeek officially released its reasoning-focused model "R3." The model outperforms OpenAI's "o3" by 0.8 points on the AMATH (Advanced Mathematics Reasoning) benchmark and sets a new record of 72.4% on SWE-bench Verified. Model weights were simultaneously released under the MIT License, with API pricing approximately one-eighth that of o3. It marks the first time the three conditions of "open, affordable, and benchmark-leading" have converged simultaneously, and this week marks the beginning of a fundamental reassessment of how enterprises approach reasoning AI adoption.
At 17:00 UTC on August 29, 2026, DeepSeek simultaneously published the model weights and technical report for "DeepSeek-R3" on Hugging Face. The total parameter count is 671B (Mixture-of-Experts architecture), but active parameters during inference are kept to 37B, providing a significant cost advantage over dense models of comparable scale.
Key benchmark results (from the official technical report):
On X (formerly Twitter), developer verification posts flooded in immediately after release, with "R3 AMATH" trending in Japan late on the night of the 29th.
"Tried DeepSeek R3. If the SWE-bench score is real, we need to revisit our backbone model selection for coding agents before the week is out. Running internal comparative evaluations right now."
DeepSeek released "R1" in January 2025, and strengthened multi-step reasoning (extended Chain-of-Thought) with "R2" in July of the same year. R3 is the successor to R2, arriving approximately 13 months later. The company is headquartered in Beijing, with researchers from Tsinghua University and Peking University forming its core team.
The MoE architecture compresses computational costs during inference from the 671B scale down to the equivalent of 37B, a mechanism also adopted by Google's Gemini series and Mistral's "Mixtral." R3 is said to further improve the efficiency of activation selection, though independent verification is still ongoing.
API access launched on the same day. The listed price is $0.55 per 1M input tokens — approximately one-eighth the level of o3 ($4.40).
The MIT License permits commercial use, modification, and redistribution without restriction. Deployment on proprietary infrastructure becomes a realistic option, and multiple engineers have already estimated that inference costs on AWS and Azure could be reduced by up to 65% compared to the o3 API. The scenario in which domestic SIers and cloud vendors fine-tune R3 for vertical SaaS applications is expected to begin materializing within the next 90 days.
SWE-bench Verified measures whether a model can automatically fix real GitHub issues. A score of 72.4% means the model can produce fixes equivalent to human pull requests in more than 7 out of 10 cases. Automated pre-processing of code reviews and automated bug triage are now coming within range.
As of the end of 2025, the conventional wisdom held that "US players lead on frontier models." R3's top ranking on AMATH serves as a concrete counterexample to that narrative. However, independent verification of reproducibility remains limited, and in the past there have been cases where announced figures and third-party verified figures diverged by 5–8 points. It is too early to draw firm conclusions.
Risk assessment when integrating models from Chinese companies into enterprise systems is a mandatory step in Japan's financial, medical, and public sectors. The fact that risk profiles differ significantly between API use and on-premises deployment is an easily overlooked point.
What sets R3 apart from other releases is that the three elements of "open weights, disruptive pricing, and benchmark leadership" have aligned simultaneously. Normally, any one of these alone would be sufficient news, but when all three converge, the market adjustment pressure increases multiplicatively. The memory of Nvidia's market capitalization dropping by approximately $590 billion in a single day following the R1 release in 2025 is still fresh.
That said, rushing to adoption decisions is risky. The immediate priority should be to "launch an internal comparative evaluation process against your own use cases before the week is out." In particular, the three areas of coding agents, legal document analysis, and mathematical optimization are high priority. Conversely, for customer-facing conversational tasks where response speed and stability take precedence over reasoning accuracy, a high benchmark score does not automatically translate into an advantage — this too warrants caution.
Multiple independent replication experiment results are expected to be published within the coming week, and assessments are likely to shift significantly once those emerge.
DeepSeek R3 may become the first model to fulfill the combined conditions of "open × affordable × high-performance reasoning." The period around next week — when third-party benchmark verification and confirmation of operational stability converge — will be the first meaningful decision-making window. When has your organization set the next trigger for its model evaluation cycle?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
