Command R (08-2024) Guide: Cheap Cited RAG From a 2024 Model

Command R 08-2024 is two years old, which in this field normally means retired. It is still listed, still supported by Cohere, and still one of the cheaper ways to get answers that cite the documents they came from. The question is whether that is enough in 2026, when Cohere's own newer models cost only about twice as much.

This guide covers what the model is good at, why its dated name matters more than it looks, and a straightforward rule for when to keep it and when to move.

What the August 2024 refresh changed

Cohere shipped command-r-08-2024 as an update to the original Command R. The open-weights release is a 32B dense transformer with grouped-query attention and a 128K context window, trained on 23 languages and evaluated on ten: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic and Simplified Chinese.

Cohere's changelog listed the improvements over the previous Command R:

• Around 50 per cent higher throughput and 20 per cent lower latency, with the serving hardware footprint cut in half.
• Better maths, code and reasoning, described as competitive with the previous version of the larger Command R+.
• Better tool-use decisions, including whether to call a tool at all.
• Better multilingual RAG, searching and answering in the user's language.
• Declining questions the supplied documents cannot answer, and better citation quality.
• Notable gains in length and formatting control.

The "declines unanswerable questions" item is the one to underline for RAG. A model that says "the documents do not cover this" is far more useful than one that fills the gap from its training data.

A dated snapshot, and why that matters

The 08-2024 suffix is not decoration. It identifies a fixed snapshot whose behaviour should not change under you. The unsuffixed alias command-r pointed at the earlier 03-2024 release, and Cohere deprecated that model in September 2025, naming command-r-08-2024 as one of the recommended replacements.

That is the pattern to design for. Anyone calling the alias got a model change they did not choose; anyone pinned to a dated id got a deprecation notice and a migration window. Pinning model versions covers why the second is the only sane default for production. As of Cohere's current deprecation page, command-r-08-2024 is not scheduled for retirement, but a two-year-old snapshot will not last indefinitely.

Citations and safety modes

The model is trained for grounded generation. You supply document snippets, ideally 100 to 400 words each, and it returns inline citations pointing back to them. There are two citation modes: accurate, which first predicts which documents are relevant and which it will cite, and fast, which answers with citations directly and spends fewer tokens. For high-volume pipelines, fast mode is the obvious place to save money; measure whether citation precision drops on your data first.

The refresh also introduced a safety_mode parameter on Cohere's API, in beta at the time. STRICT avoids sensitive topics entirely and suits customer-facing enterprise use; CONTEXTUAL, the default, allows wider-ranging output while still rejecting harmful requests; NONE opts out. Core harms remain blocked regardless. If you are serving a narrow support bot, setting STRICT is a cheap extra layer.

Tool calling and structured outputs are both supported on OpenRouter's listing. The open-weights card notes it is tuned for code explanations and rewrites, not pure code completion.

The cost case in 2026

Cohere priced the refreshed model at $0.15 per million input tokens and $0.60 per million output, and OpenRouter lists the same figures at the time of writing. Maximum output is listed at 4,000 tokens, which is fine for cited answers and too short for long documents.

Now compare with Command A+, which OpenRouter lists at $0.30 and $1.50. That is twice the input price and two and a half times the output price, for a 2026 model with vision, reasoning, 48 languages, a much larger output budget and an Apache 2.0 licence. On most RAG workloads input tokens dominate the bill, so the real gap is closer to two times than to the output ratio.

Two times is not nothing at very high volume. But it is a small enough gap that Command R 08-2024 now wins mainly where answers are short, retrieval is good, and the questions are simple enough that a 2024 model's reasoning is not the bottleneck.

Command R or its bigger sibling

Command R+ 08-2024 is the larger model from the same refresh, at $2.50 and $10 per million tokens: more than sixteen times the price. For single-step cited answers, Command R was always the better value. R+ earned its price on harder multi-step tool use and reasoning, and in 2026 those jobs are better served by newer models than by either.

If you want something even smaller from Cohere, Command R7B is the lightweight option in the same line.

When to keep it, and when to move

1. Keep it if you have a working, evaluated RAG pipeline on it, answers are short, and volume is high enough that a doubling in cost matters.
2. Move to Command A+ if you need images, longer outputs, more languages, reasoning, or self-hosting under a commercial licence.
3. Start new projects elsewhere. Building fresh on a 2024 snapshot means scheduling a migration you could have avoided.
4. Either way, run your evaluation set on both models with citation precision as the headline metric before deciding.