How to Switch Between AI Models to Keep Costs Down
Claude Haiku costs $1 per million input tokens. Claude Opus costs $5. Sonnet sits in between at $3. Using the right model for the right task can cut your AI bill by 70-80% with no meaningful quality loss.
Current Anthropic model pricing (March 2026)
Anthropic charges per token โ roughly 750 words = 1,000 tokens. Input (what you send) and output (what the AI replies) are priced separately. Prices below are per million tokens, from the official Anthropic pricing page:
| Model | Input (per MTok) | Output (per MTok) | Best for |
|---|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 | Budget tasks |
| Claude Haiku 3.5 | $0.80 | $4.00 | Cheapest option |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Most daily tasks |
| Claude Opus 4.6 | $5.00 | $25.00 | Complex reasoning |
For context: a typical conversation of 2,000 input tokens and 500 output tokens costs roughly $0.004 on Haiku 4.5 versus $0.013 on Sonnet 4.6. That's 3x more expensive for the same conversation. Across hundreds of daily automated tasks, this adds up quickly.
What tasks need which model
Think of it like hiring: you wouldn't pay a senior consultant to sort your inbox. The same logic applies here.
- Haiku (cheapest): Sorting and classifying emails, summarising long documents into bullet points, answering simple factual questions, routing tasks to the right person, sending scheduled reports
- Sonnet (mid-range): Writing proposals and emails, market research and competitive analysis, drafting content, analysing spreadsheet data, most day-to-day creative work
- Opus (premium): Complex multi-step reasoning, nuanced strategic decisions, code architecture, situations where being wrong has significant consequences
How OpenClaw makes switching practical
Without a proper setup, most people use one model for everything. OpenClaw lets you assign different models to different tasks and agents. Your background automation (monitoring news, scanning emails, generating daily reports) runs on Haiku. Your direct conversations run on Sonnet. You reach for Opus only when the task genuinely warrants it.
Switching models in OpenClaw is a configuration change, not a rebuild. You can change the model for any cron job, agent, or session independently.
๐ก Real numbers: Running nine automated cron jobs on Claude Haiku instead of Opus reduces that part of the cost by roughly 80%. Our own setup went from an estimated โฌ40-60/month to under โฌ15/month for the automated tasks after switching scheduled jobs to Haiku.
OpenAI and other providers
OpenClaw also supports OpenAI's models. GPT-4o mini is competitive with Haiku in the budget tier (around $0.15 per million input tokens as of early 2026 โ significantly cheaper than Claude Haiku). For specific tasks, mixing providers makes sense. The key advantage of OpenClaw is that you're not locked in: swapping providers is a one-line change.
The subscription alternative
Claude Pro (subscription) costs โฌ22/month (as of March 2026) and gives you access to all models with usage limits. For light, conversational use, this is often the right choice. The usage-based API model makes more sense when you're running automation with consistent, predictable volumes โ or when you want to use multiple providers strategically.
Sources
Want a cost-optimised setup from day one?
We build it with cost efficiency built in โ the right model for every task.
Book a free call โ