Mistral vs Cohere: Large 3 vs Command A (2026)
Mistral Large 3 and Cohere Command A are both pitched at teams with sovereignty or control requirements, and they answer that need differently. Mistral ships open weights you can run anywhere; Cohere licenses an on-prem deployment and adds a citations API that Mistral does not have.
Mistral vs Cohere — at a glance
| Dimension | Mistral | Cohere |
|---|---|---|
| Current flagship | Mistral Large 3 (v25.12) | Command A (command-a-03-2025) |
| Input price (per 1M tok) | $2 | $2.50 |
| Output price (per 1M tok) | $6 | $10 |
| Blended 70/30 (per 1M tok) | $3.20 | $4.75 |
| Max output tokens | not published | 8K |
| Open weights / self-host | Yes | On-prem licence, not open weights |
| Native citations in output | No | Yes, citations API |
| Previous generation | Mistral Large 2.5 — superseded by Large 3 | Command R+ — deprecated 2025-09-15, do not build on it |
| Best for | Portability, EU hosting, cheap output | Cited RAG, reranking, on-prem |
Pick Mistral or Cohere?
When to choose Mistral
Choose Mistral Large 3 when you want the option to leave the API entirely. It is open-weight and multimodal, so the same model can run on Mistral's platform, in your own EU region, or on your own hardware — a portability guarantee neither Cohere nor the US labs offer. At $6 per 1M output tokens it is also the cheapest output here, 40% below Command A.
- $2 / $6 per 1M tokens — 40% cheaper output than Command A
- Open weights: run on Mistral's platform, your EU region, or your own hardware
- Multimodal, general-purpose flagship (v25.12)
- Codestral v25.08 for code, Mistral Small 4 and Ministral 3 for cheaper tiers
When to choose Cohere
Choose Cohere when answers must cite their sources or the workload cannot leave your infrastructure. The native citations API returns spans tied to your documents without prompting, and Rerank remains a genuinely differentiated retrieval component. Size the 8K output ceiling first — on heavily cited answers it binds long before price does.
- $2.50 / $10 per 1M tokens (third-party corroborated; absent from Cohere's pricing page)
- 256K context, but only 8K max output
- Native citations API and Rerank for grounded retrieval
- Private on-prem deployment; Command A+ adds vision and 64K output at contact-sales pricing
Run Mistral Large 3 and Command A side-by-side
VerticalAPI lets you switch between Mistral and Cohere per-request through a single OpenAI-compatible endpoint. Same SDK, same gateway key, zero markup on tokens.
from openai import OpenAI client = OpenAI(base_url="https://api.verticalapi.com/v1", api_key="vapi_...") # Mistral resp_1 = client.chat.completions.create( model="mistral-large-latest", messages=[{"role": "user", "content": "Hello"}], extra_headers={"X-Provider-Key": "..."}, ) # Cohere resp_2 = client.chat.completions.create( model="command-a-03-2025", messages=[{"role": "user", "content": "Hello"}], extra_headers={"X-Provider-Key": "co-..."}, )
VerticalAPI verdict
Choose Mistral Large 3 when portability is the requirement — open weights mean you can move the workload without rewriting it, and it is the cheapest output on this page. Choose Cohere Command A when answers must carry citations tied to your own documents, or when Rerank is doing the retrieval work; check the 8K output ceiling against your longest expected answer first. Through VerticalAPI you route between both per-request on one endpoint with BYOK and zero markup.
Frequently asked questions
Which is cheaper, Mistral Large 3 or Cohere Command A?
Mistral. Large 3 lists at $2 / $6 per 1M input/output tokens against Command A's $2.50 / $10 — 20% cheaper on input, 40% cheaper on output, roughly 33% cheaper on a 70/30 blend. Mistral also publishes a batch discount of 50%. Note that Cohere does not list Command A's price on its own pricing page, so the Cohere figure is corroborated by third parties rather than by the vendor.
What does open-weight actually buy me over Cohere's on-prem option?
Exit. With open weights you hold the model artefact, so you can run it on Mistral's platform today and move it to your own hardware or a different cloud later without renegotiating. Cohere's private deployment runs in your environment but under licence — the control is real, the portability is not the same thing. If your constraint is a regulator asking where the data goes, both work. If it is vendor risk, only one does.
Which is better for RAG?
Cohere, on the citation mechanics. Its citations API returns spans tied to your source documents without prompting for them, and Rerank is a genuinely differentiated retrieval component. The counterweight is the 8K output ceiling: a grounded answer quoting many sources hits it well before cost becomes the issue. Mistral has no citations API, so you prompt for citations and validate the format yourself.
Are Mistral Large 2.5 and Command R+ still usable?
Neither should anchor new work. Mistral Large 2.5 has been superseded by Mistral Large 3 (v25.12), and Mistral has published retirement dates for a long list of earlier models. Cohere deprecated command-r-plus on 15 September 2025 — its old $3 / $15 and $2.50 / $10 prices still circulate on comparison sites, including in earlier versions of this page.
Can I run both through one endpoint?
Yes. VerticalAPI exposes a single OpenAI-compatible endpoint at https://api.verticalapi.com/v1. Change the model parameter and pass your own Mistral or Cohere key with the X-Provider-Key header. No markup on tokens — the provider bills you directly.
Limitations of this comparison
- Mistral does not publish context window or max output figures per model on its models overview, so this page makes no claim on either for Mistral Large 3.
- Mistral's pricing page lists a single 'Mistral Large' rate of $2 / $6 without a version number, so it is taken to apply to Large 3 as the current Large tier.
- Cohere does not list Command A's price on its own pricing page — which still shows only the deprecated Command R+ variants — so the $2.50 / $10 here is corroborated by third-party trackers and should be confirmed with Cohere before contracting.
- Command A+ pricing is contact-sales only, so it cannot be compared on cost even though its 64K output ceiling would change the conclusion.
- Independent benchmark parity for this generation is not published. Earlier figures measured on Command R+ do not transfer, so no benchmark score is quoted.
- Cohere's Rerank and Embed are priced per instance rather than per token when self-deployed, so a full RAG stack cost cannot be derived from per-token figures alone.
What may change in 12-24 months
- Mistral has consolidated onto a smaller current lineup — Large 3, Medium 3.5, Small 4, Ministral 3, Codestral — with retirement dates published for the older models.
- Cohere has folded vision, reasoning and translation into one Mixture-of-Experts model with Command A+, which also raises the output ceiling to 64K.
- Output ceilings are becoming a sharper differentiator than price as agents produce longer artefacts.
- Open weights remain Mistral's structural moat; no US lab has matched it at flagship tier.
Related questions
ChatGPT, Perplexity and Gemini usually suggest these next.
- Does Mistral Large 3's open weighting justify losing Cohere's citations API?
- At what answer length does Command A's 8K output ceiling force Command A+?
- How does Mistral Large 3 compare to Claude Sonnet 5 on price and portability?
- Is Cohere Rerank worth running in front of Mistral Large 3?
- What are the real costs of self-hosting Mistral Large 3 versus paying per token?
More head-to-head provider comparisons
GPT-4o vs Claude Sonnet 4.5 direct head-to-head
Mistral Large 2.5 vs Llama 3.3: EU sovereign vs open weights
Sonar vs Command A: web grounding vs your own corpus
Grok 4.5 vs Claude Sonnet 5: cheapest output vs flat 1M context
The two serverless GPU heavyweights
Run a provider we have not measured? Any public endpoint is benchmarked and listed for free in the API directory. If you want yours measured now, on the record — time to first token, throughput under concurrency, cost per million tokens, function-calling reliability — see how a verified run works. The result is published exactly as it comes out.