Moving my outreach workflow to HY3: roughly $0.02 per contacted company

I moved my outreach project from more expensive models to Tencent HY3 through Novita’s API. In my testing, HY3 handled this particular workflow better than the other models I tried and ended up being the cheapest option.

The workflow

1. Research the company: inspect its website, understand the business, and collect relevant context for outreach.
2. Find contact routes: collect public business email addresses and locate website contact forms.
3. Draft the message: use the research to generate a relevant outreach message for each company.
4. Submit via contact form: use Playwright and model-driven browser actions to fill in the form, submit the message, and check for confirmation. Cases requiring CAPTCHA or manual review are tracked separately.
5. Track results: store progress, drafts, submission outcomes, token usage, and estimated costs in SQLite.

Email delivery is a separate step. The prepared email queue already has collected addresses and generated messages, but those emails haven’t been sent yet.

The numbers

The HY3 token rates recorded in my runs are:

• $0.14 per million input tokens
• $0.035 per million cached input tokens
• $0.58 per million output tokens

So far:

• 4,707 companies researched
• 3,482 companies with generated drafts
• 364 companies successfully contacted via website contact forms
• $76.59 in total estimated HY3 spending

For the 364 companies contacted via contact forms, research, drafting, and submission—including retries—averaged $0.020 per company in estimated model costs.

Dividing the entire $76.59 spend by those 364 completed submissions gives $0.21 per successful contact so far. That broader figure includes unsuccessful attempts and preparation already completed for future outreach.

Sending the prepared emails should require little or no additional model spending, since the addresses and message content are already available. As those emails go out, the average model cost per contacted company should decrease.

The stack is Go, SQLite, Playwright, and HY3 via Novita. For this workflow, HY3 gave me the best practical results at the lowest cost among the models I tested.

These are results from my own workflow, not a general model benchmark. Costs are estimated model costs only, excluding infrastructure, email delivery, and manual work. A successful form submission does not mean a reply or a sale.