Command A+: separate limited API access from a production quote

Published 2026-10-30 · updated 2026-11-18

Cohere’s Command A+ documentation, checked November 18, 2026, identifies the model as command-a-plus-05-2026 and describes a mixture-of-experts architecture. It says trial and production keys can use it without charge until rate limits are reached, and points production use to Model Vault.

That is a defined access policy, not an unlimited production price. The previous article’s claim of a newly priced comparison was unsupported. It also incorrectly extrapolated a dense architecture from another model and described an unverified customer incident. Those sections have been removed.

Get the production terms in writing

Ask for the exact deployment option, region, capacity, billing unit, minimum commitment, overage rules and support terms. Confirm what happens when the stated limit is reached: a rejected request, a required capacity upgrade and an automatically billed overage are different outcomes.

A rate-limited free API and a reserved production deployment are not interchangeable offers. Neither a missing row in our model catalog nor another provider’s price establishes Cohere’s own contract terms. Verify the precise endpoint and version attached to any intermediary offer.

Compare costs using the actual billing model

For a metered text API, calculate input millions times the input rate plus billed output millions times the output rate, with cache and other categories accounted for separately. For reserved capacity, use committed capacity times billable hours plus support, storage, networking and other quoted charges.

As a hypothetical example, a reserved deployment costing 900 USD per month has the same direct bill as 300 million tokens at a blended metered rate of 3 USD per million. That blended rate is valid only for the assumed fixed input/output mix. Neither number is a Command A+ quote, and the equality says nothing about whether the deployment can serve the workload.

Keep task quality, concurrency, response deadlines, residency and operational responsibility alongside the cost sheet. Test accepted-task rate and retries with the same evaluation set before comparing different models. Migration and maintenance can outweigh a small difference in token charges.

Do not infer hardware from an analog

Another model’s parameter count does not specify Command A+’s weights, license or supported deployment. This guide does not establish whether a downloadable checkpoint is available or certify a self-host configuration. Obtain explicit vendor evidence for those decisions.

Use GPU offers only for a supported alternative you have validated. The cost crossover worksheet explains the additional capacity check. If the production quote or capacity evidence is missing, record the comparison as incomplete instead of filling the gap with unrelated model prices.

Reviewed November 18, 2026. This guide now explains documented access and procurement questions; it does not promise a substantiated per-token ranking that the evidence cannot support.

Put the numbers to work. Run your own workload through the breakeven calculator with today's floors, then price hardware: rent at Vast.ai → · rent at RunPod → Rental links are referral links. They never affect our rankings, which are ordered by price alone.

← All guides