AI・テクノロジー
Cohere's "Command R+ v3" Officially Released — 41% RAG Accuracy Improvement Changes the Design Assumptions for Enterprise Document Search
機械翻訳 / Machine-translated
On August 3, 2026, Cohere officially released "Command R+ v3," a model purpose-built for enterprise RAG. Internal document search accuracy (F1 score) improved 41% over the previous version, context length expanded from 128K to 512K tokens, and supported languages grew from 29 to 47. API pricing remains unchanged, while the range of design options has widened considerably.
According to the official announcement, the key changes in Command R+ v3 fall into three areas:
API pricing remains unchanged at $2.50 per million input tokens and $10.00 per million output tokens.
Reactions from enterprise AI practitioners appeared quickly on X.
"Tried Command R+ v3 on our internal knowledge base — the error rate dropped to less than half. The citation accuracy on long regulatory documents is on a completely different level. We're going to need to rethink our pipeline design."
— Enterprise AI engineer (X, approx. 3,800 followers)
While GPT-4o and Claude compete on generality, Cohere has consistently positioned "enterprise RAG specialization" as its strategic axis since 2023, focusing narrowly on accuracy improvements for vertical use cases such as legal documents, financial materials, and technical manuals.
Command R+ is its flagship model, and adoption has expanded primarily among Fortune 500 companies from the initial 2024 release through the 2025 v2. With v3, the design philosophy has shifted toward "feeding all internal documents without chunking," reflecting a change at the architectural level.
512K tokens corresponds to roughly 380,000 Japanese characters. That is enough to fit an entire set of corporate work rules and internal policy manuals, or the full text of a quarterly earnings report, into a single prompt. The conventional standard architecture of "chunk documents → vector DB search → inject top-k results into the prompt" can now realistically shift to "feed the full document as-is."
On the Japanese Document QA benchmark published by Cohere, the score improved from 62.4 in v2 to 79.1 in v3. This marks the first time a non-domestic specialized RAG model has surpassed GPT-4o's score of 74.8. Japanese has long been an area where repurposing English-optimized models hits its limits, due to the language's phrase structure, honorifics, and abbreviation conventions. Practical assessments in Japan's finance, manufacturing, and legal sectors may now shift.
Traditional RAG has been designed around balancing "chunking cost + index management cost + retrieval latency" against "accuracy." If 512K can be used cost-effectively, a design that accepts higher inference cost in exchange for eliminating index management becomes viable. Which approach is optimal depends on document volume, query frequency, and accuracy requirements. Rather than a straightforward "switch to v3," a use-case-by-use-case reassessment is needed.
The "accuracy ceiling" of RAG has long been thought to reside on the retriever side — the vector DB and search logic. What v3 demonstrates is the simple fact that raising a model's contextual comprehension also raises the upper bound of accuracy.
The improvement margin in the Japanese F1 score (+16.7 points) likely reflects not just fine-tuning, but deliberate decisions in training data composition and architecture. Cohere has not disclosed the details of its training data, but benchmark figures serve as a practical basis for business decisions.
The contrast with competitors' moves also matters. While Perplexity's "Sonar Pro 2" differentiates on real-time search, Cohere is carving out its niche by deepening accuracy on static documents. The use-case divide of "real-time external information vs. closed internal documents" is likely to define the enterprise RAG market landscape in the second half of 2026.
Who moves next — Microsoft with an update to its Azure AI Search integration, or Google with improvements to grounding accuracy? Now that Cohere's v3 has raised the industry's "accuracy benchmark," a competitive response will likely come quickly.
Command R+ v3 has shifted the premise that "RAG is a cost-versus-accuracy tradeoff" by one step. The improvements to the 512K context and Japanese accuracy give organizations in Japan's legal, financial, and manufacturing sectors an opportunity to reassess RAG adoption. After taking stock of your organization's document volume and query characteristics, now may be the right time to revisit your pipeline design.
Is your organization's RAG pipeline still running on the assumption that chunking is required?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
