Cohere’s public pricing page is built for enterprise buyers. North and Compass are sold through a contact form. Model Vault, the dedicated inference platform, is priced per instance hour or per month. Current self-serve API rates are not listed there: the page’s FAQ quotes prices for older Command models instead.
That is a problem for developers who want to estimate a bill before a sales call. The five alternatives below publish per-million-token prices, which makes a cost model possible in a spreadsheet.
The comparison covers the three jobs Cohere is usually bought for: text generation with Command, embeddings with Embed, and reranking with Rerank. Keep those jobs separate when you read the table, because no single provider wins all three.
Why do teams look for Cohere alternatives?
The first reason is pricing transparency. A developer who wants to estimate next month’s bill cannot do it from Cohere’s pricing page alone. North and Compass require a call, and current Command API rates are not listed there.
The second is commitment. Model Vault is sold per instance hour or per month, with a $2,000 monthly floor for the smallest Embed configuration. That suits steady production load and punishes experimentation.
The third is trial restrictions. Cohere’s own FAQ says trial keys are rate limited and may not be used for production or commercial purposes. Teams that prototype on a trial key must move to a production key before launch.
The fourth is model fit. Command models are strong at retrieval-augmented generation and enterprise search, but teams building coding agents often prefer a different family entirely.
Embeddings and reranking: the real price gap
Cohere’s Model Vault lists Embed 5 at $3.00 per hour for a small instance, or $2,000 per month. Voyage AI lists voyage-4 at $0.06 per million tokens, with the first 200 million tokens free. Those units are not directly comparable, but the difference in commitment is obvious.
For a team indexing ten million tokens, Voyage’s published rate implies a bill of $0.60 once the free allowance is used. A dedicated instance costs $2,000 per month whether you use it or not. Self-serve token pricing wins for bursty workloads, and dedicated instances win for steady high volume with strict isolation.
Reranking follows the same pattern. Voyage rerank-3 costs $0.05 per million tokens and the first 200 million are free. Cohere’s Rerank 5 is sold as a $4.00 hourly instance or $2,500 monthly. Again, the choice depends on utilisation, not on a single headline number.
Language models: what a Command migration costs
Cohere’s FAQ still publishes rates for older Command models. Command is listed at $1.00 input and $2.00 output per million tokens. Command R is $0.50 and $1.50. Command R+ 08-2024 is $2.50 and $10.00. Aya Expanse is $0.50 and $1.50.
Against those rates, OpenAI’s gpt-6.1-terra at $1.00 and $5.00 is more expensive on output and equal on input. AI21’s Jamba Mini at $0.20 and $0.40 is far cheaper. Anthropic’s Sonnet 5.5 at $2.00 and $10.00 sits above the legacy Command R tier.
Cheaper output tokens are not automatically cheaper answers. A model that needs two attempts, or a longer prompt, can cost more than a stronger model that answers once. Measure cost per completed task, not cost per token.
What breaks when you switch providers?
Embeddings are not portable. Vectors from different models live in different spaces, so switching means re-indexing the entire corpus. Check dimensions before you migrate: Perplexity’s 0.6B model returns 1024 dimensions and its 4B model returns 2560, while Voyage publishes dimensions per model.
Rerank APIs also differ. Some accept a query and a list of documents, others accept pre-tokenised pairs. Response fields and score ranges are not standardised. Budget a week of integration work even for a small pipeline.
Finally, check rate limits and trial terms. A production key behaves differently from a trial key at every vendor on this list, and the published price usually assumes production access.