DeepSeek + Qwen in Claude Code: One Key, Many Models
TL;DR — You don't need separate accounts to mix models. With a single sk-relay-... key, the same tooling you already point at Claude can also reach DeepSeek and Qwen. Running deepseek claude code workflows is just a matter of changing the model string in your request — the base URL, the key, and the wallet stay the same. Use the cheaper open and Chinese models for bulk or simple work, and keep Claude for the hard turns. This guide shows how the routing works, what the listed prices look like, and when to reach for which model.
Why one key for DeepSeek, Qwen, and Claude
The usual multi-model setup is a mess: a DeepSeek account here, a Qwen key there, a Claude subscription somewhere else, three dashboards, three invoices, three places to top up. The relay collapses that. llmapi.pro is a Claude-compatible API relay that exposes additional models beyond Claude on the same wallet and the same key. The live roster lives at /all-models — that's the source of truth, not this post.
What that buys you in practice:
- One balance to fund, one key to rotate, one place to watch spend.
- The Anthropic Messages protocol you already speak (
/v1/messages) plus the OpenAI protocol (/v1/chat/completions,/v1/responses). - No second SDK, no second auth flow. You change a model ID and that's it.
The catch — and we say this in every post because we'd rather you trust us — a relay is a third party in your request path. If you need a direct contractual relationship with a model vendor, or you're contractually barred from routing through an intermediary, that's a real reason not to use a relay. For most cost-sensitive and multi-model developers, the single-wallet convenience wins.
How to use deepseek claude code (and Qwen) on the same setup
Claude Code points at the relay through the two documented environment variables. Once those are set, every request from Claude Code flows through your relay key:
export ANTHROPIC_BASE_URL=https://llmapi.pro
export ANTHROPIC_API_KEY=sk-relay-your-key-here
From there, the model you talk to is decided by the model string. Claude IDs route to Claude. Open and Chinese model IDs route to those models. You can confirm what's currently advertised with a plain GET:
curl https://llmapi.pro/v1/models \
-H "Authorization: Bearer sk-relay-your-key-here"
If you drive the relay directly over the OpenAI protocol — handy for scripts, batch jobs, or a second tool that isn't Claude Code — the model field is where you pick DeepSeek or Qwen:
curl https://llmapi.pro/v1/chat/completions \
-H "Authorization: Bearer sk-relay-your-key-here" \
-H "Content-Type: application/json" \
-d '{
"model": "<deepseek-model-id-from-/all-models>",
"messages": [{"role": "user", "content": "Summarize this diff in one line."}]
}'
Check the exact model IDs on /all-models before you hardcode them — names and versions move, and we'd rather you copy the live string than a stale one from a blog.
For a full step-by-step Claude Code + open-model walkthrough, the Claude Code with Qwen setup guide covers the env-var and config details end to end.
When each model makes sense
The whole point of one key is that you stop treating model choice as a commitment and start treating it as a per-task decision. A pattern our multi-model users settle into:
| Task | Reach for | Why |
|---|---|---|
| Bulk transforms, boilerplate, log triage, simple edits | DeepSeek / Qwen | Cheap, fast enough, the answer space is narrow |
| Long-context reads, routine summaries, first-draft scaffolding | Qwen | Solid on structured, repetitive work |
| Architecture, gnarly debugging, the turn that has to be right | Claude (Opus/Sonnet) | You pay more, but a wrong answer here costs more than the tokens |
The spend math is the lever. Here are example listed prices — always verify the live numbers on /pricing and /all-models, because they change:
| Model | Listed price (RMB / million tokens) |
|---|---|
| DeepSeek V4 Pro | 1.74 in / 3.48 out |
| Qwen 3.6-27B | 0.3 in / 3.2 out |
| FLUX.1-dev (image) | 0.014 per image |
The gap between an open model's input price and a frontier Claude turn is wide enough that routing the easy 80% of a session away from Claude noticeably changes your bill — without making the hard turns any worse, because those still go to Claude. If you want to go deeper on that strategy, see Claude Code cost optimization for 2026, and for a head-to-head on coding quality, Qwen Coder vs Claude Sonnet for coding.
The wallet: how billing actually behaves
Two things matter here and both are in your favor:
- Balance never expires. Top up once, draw it down across DeepSeek, Qwen, Claude, and image models. There's no monthly reset clock forcing you to spend or lose it.
- Recharge tiers carry volume bonuses. Larger top-ups get credited extra:
| Recharge | Bonus |
|---|---|
| 100 RMB | +5% |
| 500 RMB | +10% |
| 2000 RMB | +15% |
So a 2000 RMB top-up lands as 2300 RMB of usable balance, and that balance is fungible across every model on the roster. Because it doesn't expire, a bigger one-time top-up is the rational move if you're going to use the relay at all — you're not betting on a usage forecast.
A realistic multi-model session
Here's what "one key, many models" looks like in a single afternoon of work, with no account switching:
- Triage — point a cheap Qwen call at fifty stack traces to cluster them. Pennies.
- Bulk edit — DeepSeek rewrites a hundred boilerplate test stubs. Still cheap.
- The hard turn — the one race condition that's actually subtle goes to Claude Opus or Sonnet, where being right matters more than the token cost.
- Asset — need a placeholder graphic? A FLUX.1-dev image call comes off the same wallet.
Every one of those billed against the same balance, authenticated with the same sk-relay-... key, observed in the same spend view. That's the entire pitch: the friction of using the right model for each task drops to changing a string.
When NOT to do this
Be honest with yourself about the fit:
- You only ever use Claude. Then the multi-model angle is irrelevant — you just want a Claude-compatible endpoint, and any of our Claude-focused guides cover that.
- You need a direct vendor relationship for compliance, SLA, or contractual reasons. A relay is an intermediary; that's a deliberate trade.
- Your task is latency-critical and single-model. Adding a hop you don't need buys you nothing.
For everyone else — the developers juggling cost against capability and tired of three logins — one wallet that reaches Claude plus open and Chinese models is the simpler shape. Grab a key at /register, set the two env vars, and pick your model per task instead of per account.
llmapi.pro is an independent, Claude-compatible API relay; we are not affiliated with Anthropic. DeepSeek, Qwen, FLUX, and Claude are used for identification only.