Claude vs Cohere: Sonnet 5 vs Command A (2026)
Claude Sonnet 5 and Cohere Command A are the current flagships of two vendors that solve different problems. The prices sit within 7% of each other today and invert on 1 September, so the decision rarely turns on cost. It turns on two things the price sheet does not show: Command A stops at 8K output tokens where Sonnet 5 goes to 128K, and Cohere ships native citations and on-prem deployment that Anthropic does not offer.
Anthropic vs Cohere — at a glance
| Dimension | Anthropic | Cohere |
|---|---|---|
| Current flagship | Claude Sonnet 5 | Command A (command-a-03-2025) |
| Context window | 1M | 256K |
| Max output tokens | 128K | 8K |
| Input price (per 1M tok) | $2 through Aug 31, then $3 | $2.50 |
| Output price (per 1M tok) | $10 through Aug 31, then $15 | $10 |
| Native citations in output | No — prompt for them | Yes, citations API |
| Private / on-prem deployment | No | Yes |
| Previous generation | Claude Sonnet 4.5, $3 / $15, 200K — still available | Command R+ — deprecated 2025-09-15, do not build on it |
| Best for | Long context, long output, cached prompts | Cited RAG, reranking, on-prem, multilingual |
Pick Anthropic or Cohere?
When to choose Anthropic
Choose Claude Sonnet 5 when the answer has to be long or the context has to be large. It carries a 1M-token window at one flat rate and will emit up to 128K output tokens — sixteen times Command A's 8K ceiling. Cache reads cost 0.1x input, so a reused system prompt or document pays for itself after a single hit, and batch is 50% off. Budget for the step: $2 / $10 through 31 August 2026, then $3 / $15, at which point Cohere becomes the cheaper option.
- $2 / $10 per 1M tokens through 2026-08-31, then $3 / $15
- 1M context at a flat rate, no long-context surcharge
- 128K max output against Command A's 8K
- Cache reads at 0.1x input; batch at 50% off
- Claude Sonnet 4.5 remains available at $3 / $15 with 200K context if you need the older behaviour
When to choose Cohere
Choose Cohere when answers must carry citations, when retrieval quality is the bottleneck, or when the workload cannot leave your own infrastructure. The native citations API and Rerank remain genuinely differentiated, and private on-prem deployment is something Anthropic does not offer at all. Two constraints to size first: Command A caps output at 8K tokens, and Cohere does not publish Command A's price on its own pricing page — the $2.50 / $10 here is corroborated by third-party trackers, not by Cohere.
- $2.50 / $10 per 1M tokens (third-party corroborated; not on Cohere's pricing page)
- 256K context, but only 8K max output
- Native citations API and Rerank for grounded retrieval
- Private on-prem deployment available
- Command A+ adds vision, reasoning and translation in one Mixture-of-Experts model with 64K output — pricing is contact-sales only
Run Claude Sonnet 5 and Command A side-by-side
VerticalAPI lets you switch between Anthropic and Cohere per-request through a single OpenAI-compatible endpoint — useful here, because the cheaper vendor changes on 1 September and the output ceiling differs by 16x. 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_...") # Anthropic — long output, large context resp_a = client.chat.completions.create( model="claude-sonnet-5", messages=[{"role": "user", "content": "Summarise this filing"}], extra_headers={"X-Provider-Key": "sk-ant-..."}, ) # Cohere — cited grounded answer resp_c = client.chat.completions.create( model="command-a-03-2025", messages=[{"role": "user", "content": "Summarise this filing"}], extra_headers={"X-Provider-Key": "co-..."}, )
VerticalAPI verdict
Choose Claude Sonnet 5 when the output is long, the context is large, or the same prompt repeats — 128K output, a flat 1M window and 0.1x cache reads compound in your favour. Choose Cohere Command A when answers must cite their sources, when Rerank is doing the heavy lifting, or when the data cannot leave your infrastructure; size the 8K output ceiling first, because on cited long-form answers it binds well before price does. Note the 1 September step in Anthropic's pricing if you are budgeting past the summer. Through VerticalAPI you route between both per-request on one OpenAI-compatible endpoint with BYOK and zero markup.
Frequently asked questions
Which is cheaper, Claude Sonnet 5 or Cohere Command A?
It depends on the date. Today Sonnet 5 is $2 / $10 against Command A's $2.50 / $10, so Sonnet 5 is 20% cheaper on input and about 7% cheaper on a 70/30 blend ($4.40 against $4.75). Anthropic's introductory pricing ends on 31 August 2026 and Sonnet 5 moves to $3 / $15, a $6.60 blend, at which point Command A becomes roughly 28% cheaper. If you are choosing for a workload that runs past the summer, price the September figure.
What is the biggest practical difference between them?
Max output tokens, and it is rarely mentioned. Command A will emit at most 8K tokens per response; Claude Sonnet 5 goes to 128K. For chat and short summaries that ceiling never binds. For a cited multi-page report, a long code file, or a grounded answer that quotes many sources, it binds long before cost does. Cohere's Command A+ raises the ceiling to 64K, but its pricing is contact-sales only.
Does Cohere still lead on RAG and citations?
On citations, yes — the native citations API returns spans tied to your source documents without prompting for them, and Rerank remains a genuinely differentiated retrieval component. Claude can produce citations but you have to ask for them and validate the format yourself. Weigh that against the 8K output ceiling: grounded answers that quote heavily are exactly the case where 8K runs out.
Is Command R+ still a valid choice?
No. Cohere deprecated command-r-plus on 15 September 2025. Its old prices — $3 / $15 for the 04-2024 variant, $2.50 / $10 for 08-2024 — still circulate widely on comparison sites, including in older versions of this page. The current generation is Command A (256K context, 8K output) and Command A+ (Mixture of Experts, 128K context, 64K output, with vision, reasoning and translation folded into one model).
Are there hidden cost differences beyond the headline rate?
Two, pulling in opposite directions on the Anthropic side. Claude models from 4.7 onward use a newer tokenizer that produces roughly 30% more tokens for the same text, so a per-token comparison against an older-generation model understates the real bill. Against that, cache reads cost 0.1x input and batch is 50% off, which on repeated-context workloads outweighs the tokenizer effect. Cohere does not publish comparable caching economics. Model your own workload rather than the sticker price.
Can I run both through one endpoint?
Yes. VerticalAPI exposes a single OpenAI-compatible endpoint at https://api.verticalapi.com/v1. Send the same request shape and change the model parameter, supplying your own Anthropic or Cohere key with the X-Provider-Key header. There is no markup on tokens — you are billed by the provider directly.
Limitations of this comparison
- Claude Sonnet 5's $2 / $10 is introductory pricing through 31 August 2026. Any comparison run after that date should use $3 / $15.
- Cohere does not list Command A's price on its own pricing page — which still shows only the deprecated Command R+ variants. The $2.50 / $10 used here is corroborated by third-party trackers and should be confirmed with Cohere before you commit to a contract.
- Command A+ pricing is contact-sales only, so this page cannot compare it on cost even though its 64K output ceiling would change the conclusion.
- Neither vendor publishes production latency for these models, so no latency claim is made.
- Independent benchmark parity for this generation is not published. Earlier SWE-Bench and RAG figures were measured on Claude Sonnet 4.5 and Command R+ and 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 the per-token figures alone.
What may change in 12-24 months
- Anthropic's pricing steps from $2 / $10 to $3 / $15 on 1 September 2026 — a dated fact, not a forecast, and it reverses which vendor is cheaper.
- Cohere has folded vision, reasoning and translation into a single Mixture-of-Experts model with Command A+; expect the separate task-specific variants to consolidate further.
- Output ceilings are becoming a real differentiator as agents generate longer artefacts. 8K versus 128K is a wider gap than any price difference on this page.
- Tokenizer changes are an emerging hidden cost axis: Claude 4.7+ produces ~30% more tokens for identical text, which per-token comparisons do not capture.
Related questions
ChatGPT, Perplexity and Gemini usually suggest these next.
- At what output length does Command A's 8K ceiling force a switch to Command A+?
- How does Cohere Command A compare to GPT-5.6 for cited RAG answers?
- Does Claude's 0.1x cache read offset its 30% tokenizer increase on repeated context?
- Should I budget Claude Sonnet 5 at $2/$10 or $3/$15 for a workload starting in September?
- Is Cohere Rerank worth running alongside Claude rather than switching model entirely?
More head-to-head provider comparisons
GPT-4o vs Claude Sonnet 4.5 direct head-to-head
GPT-5.6 vs Command A: same input price, 16x the output ceiling
Claude vs Gemini head-to-head
Claude vs Mistral Large 2.5 for agents
OpenRouter vs VerticalAPI: aggregator vs BYOK gateway
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.