Comparison · Chat & code
DeepSeek V4 Pro vs Claude Sonnet 5
DeepSeek V4 Pro vs Claude Sonnet 5: token rates, reply length, open weights and Claude Code support. Find out which model fits your budget and your tools.
What each one accepts
| DeepSeek V4 Pro | Claude Sonnet 5 | |
|---|---|---|
| Vendor | DeepSeek | Anthropic |
| Type | Chat & code | Chat & code |
| Released | 2026-08 | 2026-06 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output | 384,000 tokens | 128,000 tokens |
| Endpoints | /v1/chat/completions | /v1/chat/completions · /v1/messages |
Token prices
| DeepSeek V4 Pro | Claude Sonnet 5 | |
|---|---|---|
| Input / 1M tokens | $0.54 | $3.24 |
| Output / 1M tokens | $1.22 | $16.20 |
| Input from cache / 1M | same as input | same as input |
| 1,000 replies, 1,500 in / 400 out | $1.30 | $11.34 |
Charged from a prepaid USD balance, no taxes added. One key runs both, through /v1/chat/completions in the OpenAI format. A failed generation is $0.
If the budget leads, go with DeepSeek V4 Pro: both its input and output rates sit below Claude Sonnet 5. If you code in Claude Code or need the /v1/messages format, Sonnet 5 is the one that plugs in. And if price is the only reason you are looking at DeepSeek, try V4 Flash too, the smaller sibling with a much lower rate than V4 Pro.
When to pick DeepSeek V4 Pro
- You feed it whole repositories or document sets and expect a short answer, so input dominates.
- You need very long output, for example a full report past 128K tokens.
- Your agent workflows were already tuned on V4 Pro.
- You want to measure an open-weight model against a closed one on your data.
When to pick Claude Sonnet 5
- You work in Claude Code every day on tests, fixes and features.
- Your agent plans and calls tools across long chains of steps, which Anthropic names as the strength of this Sonnet.
- You debug old code and need the real root cause, not a patch.
- Your code is written for the Anthropic SDK.
Differences you will notice
- Rates. DeepSeek V4 Pro is billed at $0.54 / $1.22 per 1M tokens, Claude Sonnet 5 at $3.24 / $16.20 per 1M tokens. Sonnet is the pricier one on both sides; the 1,000-reply cost above shows how much for a typical chat.
- Output length. V4 Pro can answer with up to 384,000 tokens; Sonnet stops at 128K. Both read up to 1,000,000 tokens.
- Where they plug in. Sonnet runs on
/v1/chat/completionsand/v1/messages, so Claude Code accepts it. V4 Pro uses the OpenAI-compatible endpoint only, which suits Cursor, Cline and n8n. - Weights. DeepSeek publishes the V4 Pro weights: a mixture-of-experts design of 1.6 trillion parameters, with 49 billion used for each token. Sonnet 5 is closed.
- Reasoning controls. DeepSeek gives you low, high and max effort. Sonnet 5 thinks adaptively by default and returns a 400 for non-default
temperature,top_portop_k.
How to decide
On both models output is the expensive side, so the length of answers drives the bill more than the length of prompts. Put your own token mix into the cost calculator, then run the same dozen tasks through both and read usage. My own repository has not been through this pair, so the quality verdict belongs to your trial. A ready test script is in the guide to switching LLM models in Python, and budget models are ranked in Cheapest LLM models.
Questions
Is V4 Pro going away?
DeepSeek once said it would shift V4 Pro requests onto Flash, then dropped that plan. V4.1 Pro has been mentioned without a release date, so expect to migrate one day and keep the model id in config.
Where does V4 Flash fit in?
It is the smaller V4 model and much cheaper than Pro on input and output. See V4 Flash vs V4 Pro for when the smaller one is enough.
Can I point Claude Code at DeepSeek?
Not here. Claude Code needs the Anthropic Messages endpoint, which serves Claude models only. Use DeepSeek in OpenAI-compatible tools, with the very key you use for Sonnet.