Command R+ (08-2024) Guide: A Legacy Model and How to Leave It

If you are reading about Command R+ 08-2024 in 2026, the likely reason is that something already runs on it. It was Cohere's top model until Command A arrived in March 2025, many enterprise RAG and tool-use pipelines were built and validated against it, and it still answers requests. Nobody is starting fresh on it.

So this is less a review than a migration guide: what the model is, what it costs relative to its replacements, which of its behaviours your pipeline may silently depend on, and how to move without breaking them.

What Command R+ 08-2024 is

Cohere released command-r-plus-08-2024 as a refresh of the original Command R+. The open-weights release is a dense 104B transformer with grouped-query attention, text in and text out, and a 128K context window. Like its smaller sibling it was trained on 23 languages and evaluated on ten.

Cohere's changelog described the refresh as roughly 50 per cent higher throughput and 25 per cent lower latency than the previous Command R+, on the same hardware footprint. Alongside that came the improvements shared with Command R 08-2024: better decisions about when to call a tool, better instruction following in the preamble, multilingual RAG that answers in the user's language, declining unanswerable questions, and better citations that can be switched off.

OpenRouter lists a maximum output of 4,000 tokens, with tool calling and structured outputs supported.

What it was for: multi-step tool use

The R+ model card puts multi-step tool use at the centre of its agent story. It describes an agent loop of action, observation and reflection: the model calls a tool, reads the result, decides whether it needs another call, and only then answers. Single-step tool use, where the model emits one JSON list of actions and you return the results for a final answer, is documented alongside it. If your pipeline runs the multi-step loop, that is the behaviour a replacement has to match first.

It also offered the same two citation modes as Command R. Accurate mode predicts relevant and cited documents before answering; fast mode answers with citations directly and uses fewer tokens. The model card is candid that it is tuned for code explanations and rewrites, and may not perform well at pure code completion.

For 2024, this was a well-rounded enterprise package. The problem is not that it got worse. It is that everything around it got much cheaper and much better.

Where it sits on the price list now

Cohere's pricing page now lists Command R+ 08-2024 under legacy pricing for existing customers, at $2.50 per million input tokens and $10 per million output. OpenRouter lists the same figures at the time of writing.

Put that next to the alternatives from the same vendor:

• Command A lists at the same $2.50 and $10, with a 256K window. Cohere claimed 150 per cent higher throughput than R+ 08-2024 at launch.
• Command A+ lists at $0.30 and $1.50 on OpenRouter: more than eight times cheaper on input and over six times cheaper on output, with vision, reasoning, 48 languages and Apache 2.0 weights.
• Command R 08-2024 lists at $0.15 and $0.60, for workloads where single-step cited answers are enough.

At legacy pricing, every month on R+ is a month paying several times the going rate for an older model. That makes migration a cost project as well as a quality one.

What your pipeline may depend on

The risk in migrating is not that the new model is worse. It is that your prompts, parsers and evaluation were tuned to quirks you never wrote down. Check these specifically before switching:

• Citation format. If downstream code parses citation spans, confirm the replacement emits them in the same structure, through the same API fields.
• Refusal behaviour. R+ 08-2024 was trained to decline unanswerable questions. A replacement that answers more readily can look like an improvement while quietly increasing hallucinations.
• Output length. A 4,000-token ceiling may have been masking verbose behaviour. Newer models with larger output budgets can run long and cost more than the price table suggests.
• Preamble sensitivity. Instructions tuned for one model's preamble handling often need rewording for another.
• Safety settings. If you relied on a particular safety_mode, check what the replacement's defaults are.

A migration plan that catches regressions

1. Freeze an evaluation set from production: at least fifty real requests with the documents and tool results that went with them, and the answers your users accepted.
2. Run it on R+ 08-2024 to establish a baseline under your current harness, since live behaviour may have drifted from what you validated originally.
3. Run the same set on Command A+, with reasoning off and then on, and on Command A if you need more than 128K tokens of input.
4. Score correctness, citation precision, refusal rate on deliberately unanswerable questions, and output tokens per request. Compute cost per accepted answer.
5. Shadow a slice of live traffic before cutting over, and keep R+ as a fallback until the new model has run clean for a full business cycle.

Model deprecation and migration covers the general playbook, and it is also worth pinning the replacement to a dated id, as you did with this one.

The short answer

Command R+ 08-2024 is still a working model and is not on Cohere's retirement list today. But it is priced as legacy, outclassed by its own successors, and no cheaper than any of them. Keep it only as long as it takes to validate a replacement, and for most RAG and tool-use pipelines that replacement is Command A+.