Claude Plus A Cheaper Model
Can You Use Claude as the Planner and a Cheaper Model for the Work?
Use each model for what it's best at.
Yes, you can use Anthropic's Claude as the planner and a cheaper model for the work: Claude plans and checks, while a much lower-cost model such as DeepSeek's API drafts the bulk of it. The mixed setup takes one add-on, and a full swap instead drops Claude's judgment at every step.
Published
Key Takeaways
- Anthropic's Claude Pro is US$20 a month, or US$17 a month billed annually, and already includes Claude Code.
- DeepSeek's API charges by the token: its deepseek-flash model runs US$0.30 per 1M input tokens at peak, US$0.15 off-peak, with cached input far cheaper still.
- The recommended setup keeps Claude as the planner and adds one piece, an MCP server that hands drafting tasks to the cheaper model.
- A full swap instead points every model in Claude Code at DeepSeek, which drops Claude's judgment at every step, not just the drafting.
- Never send client, customer, or company-confidential data to a model your employer hasn't approved, and have Claude check the cheaper model's drafts before you use them.
The full guide
Tap any section to open it.
About this guide
A widely shared post this week claimed you could run Anthropic's Claude on its US$20 plan as the planner while a much cheaper model does the bulk of the work for pennies. That's close to true, with real tradeoffs worth knowing before you wire it up.
What this costs, as of November 5, 2026
| Anthropic's Claude Pro | DeepSeek's API (deepseek-flash) | |
|---|---|---|
| Plan | US$20/month, or US$17/month billed annually | No plan; pay per token |
| Input tokens | Included in the plan | US$0.30 per 1M (peak), US$0.15 per 1M (off-peak) |
| Cached input tokens | Included in the plan | US$0.006 per 1M (peak), US$0.003 per 1M (off-peak) |
| Output tokens | Included in the plan | US$1.20 per 1M (peak), US$0.60 per 1M (off-peak) |
| Data location | Anthropic | Processed and stored in the People's Republic of China, per DeepSeek's privacy policy |
The IdeaWhat an orchestrator actually does
An orchestrator is just the model that plans the job and checks the result, instead of doing every step itself. DeepSeek API pricing is what makes the worker side of this so cheap.
The orchestrator plans
Anthropic's Claude reads the task, breaks it into steps, and decides what needs doing next, the same way it does today.
The worker drafts
A cheaper model, such as DeepSeek's API, writes the first pass of code, text, or research at a fraction of the cost per token.
The orchestrator checks
Claude reviews what the cheaper model produced before any of it reaches your real work, catching the mistakes a cheap model is more likely to make.
Two Ways To Set It UpPlanner plus worker, or a full swap
The recommended way keeps Anthropic's Claude as the planner on your existing plan and adds one piece: an MCP server that hands drafting tasks to a cheaper model's API, such as DeepSeek's. Claude still decides what to do and checks the result; only the drafting moves to the cheaper model.
The other way is a full swap. DeepSeek's own guide for Claude Code points every model slot at its Anthropic-format endpoint instead of Claude, so DeepSeek plans too, not just drafts. That's cheaper, but you lose Claude's judgment on every step, not only the drafting.
The Worked ExampleWhat one user actually paid
One user reported a DeepSeek dashboard showing 142 requests and 7,858,370 tokens over 30 days, for US$0.55 total. That's a strikingly small bill for nearly 8 million tokens.
The math doesn't add up at DeepSeek's listed peak rate: 7.86 million tokens at US$0.30 per 1 million peak input tokens comes to about US$2.36. The gap points to most of those tokens being cached or sent off-peak, when DeepSeek's prices are half of peak, outside its 01:00 to 04:00 and 06:00 to 10:00 UTC weekday peak window.
Heavier tasks can use DeepSeek's deepseek-v4-pro model instead, at US$1.32 per 1M peak input tokens and US$3.96 per 1M peak output tokens, still a fraction of what equivalent work would cost through most other paid APIs.
Keep This Off A Cheap WorkerWhat not to send to the cheaper model
DeepSeek's own privacy policy states that personal data is collected, processed, and stored in the People's Republic of China. That alone is a reason to pause before sending anything sensitive through its API.
Never send client, customer, or company-confidential data to a model your employer hasn't approved, no matter how cheap it is per token. Save the cheaper model for drafts you'd be comfortable showing a stranger.
Before You Use Anything It WritesHave the planner check the worker's drafts
A cheaper model is more likely to make mistakes than Claude is, which is exactly why it stays the worker and not the planner. Treat every draft it hands back as a first pass, not a finished one.
Have Anthropic's Claude read the draft before you use it: ask it to check facts, tone, and anything that touches a real person or a real decision. That one extra step is most of what keeps this setup safe to use.
Start Here
Try the first three lessons free
No account needed: see how Claude actually plans and checks work before you wire up anything else.
Learn the planner before you add a cheaper worker
Knowing when to trust a draft and when to send it back is the skill that makes this setup safe. Start with the free lessons, then see what a full agent build looks like in the AI agent kit.
Frequently Asked Questions
Can I use Claude as the planner and a cheaper model for the actual work?
Yes. Anthropic's Claude can stay the planner on your existing plan, checking and directing the work, while a much cheaper model drafts the bulk of it. The mixed setup needs one add-on that hands tasks to the other model.
What does it cost to use Claude with DeepSeek?
Anthropic's Claude Pro is US$20 a month, or US$17 a month billed annually, and already includes Claude Code. DeepSeek's API charges separately by the token: its deepseek-flash model runs US$0.30 per 1M input tokens at peak and US$0.15 off-peak, with cached input far cheaper.
How do I set up Claude as an orchestrator with a cheaper model?
An orchestrator is the model that plans the work and checks the result; the worker is the model that drafts it. People wire this up with an MCP server that hands tasks from Claude to the cheaper model's API, added as one extra connection inside Claude Code.
Is DeepSeek safe to use for work tasks?
Check with your employer first. DeepSeek's own privacy policy states that personal data is collected, processed, and stored in the People's Republic of China, so client, customer, or confidential company data should stay off it unless your employer has approved that.
What is an AI orchestrator?
An orchestrator is the model you talk to that breaks a job into smaller tasks, hands the drafting work to another model, and checks what comes back before you use it. Claude can fill that role while a cheaper model does the drafting.
Can DeepSeek replace Claude Code entirely?
Yes, but that's a full swap, not a mix. DeepSeek's own Claude Code guide points every model slot in Claude Code at its Anthropic-format endpoint instead, which drops Claude's judgment at every step rather than keeping it as the planner.
How much cheaper is DeepSeek than Claude?
It depends on usage and caching. One user reported 142 requests and 7,858,370 tokens over 30 days for US$0.55 on deepseek-flash. Our own math on that many tokens at the peak input rate comes to about US$2.36, so the reported total implies mostly cached or off-peak tokens.
What should I keep off a cheap AI model?
Keep client, customer, and confidential company information off any model your employer hasn't approved, cheap or not. DeepSeek's privacy policy states data is processed and stored in China, which may matter for your employer's data rules before you ever get to cost.
Keep Claude as the planner. Let it check everything.
The cheap model drafts. Claude decides what's good enough to use.
Nightschool AI is independent and not affiliated with Anthropic or DeepSeek; figures above are as published by each company, stamped as of November 5, 2026.