Case study · August 11, 2026
My AI read my 25 unread emails, drafted the 11 that needed replies, and told me the other 14 don't need me. Cost: 1.2 cents.
25 morning emails triaged, 11 usable drafts written, 14 dismissed as spam / already handled / FYI. Four minutes, 1.2 cents via Reality Router — same tokens would have been $0.52 on Claude Opus 5.
Twenty-five emails sat in my inbox on a Tuesday morning. The usual mix: a designer following up on mockups, mom asking about Sunday dinner, an investor cold-DMing, a data-breach class-action asking me to file a claim, a legal MSA that needed me to actually read section 7.3 before signing, an Uber receipt, a LinkedIn digest, three obvious spam messages, and one Nigerian prince.
Reading each one and deciding what to do about it: twenty to thirty minutes of low-value attention I don't have.
So I sent them all to an agent through Reality Router. Four minutes later I had a triage report. Every email categorized, priority-tagged, and — for the eleven that needed a reply — an actual draft ready to send with two clicks. Total cost: $0.012.
What the agent produced
Twenty-five emails came in. Out came a clean breakdown:
| Category | Count |
|---|---|
needs_reply (draft attached) | 11 |
spam_or_marketing (safe to ignore) | 5 |
already_handled (Stripe payout, Uber receipt, security alert I recognized) | 5 |
calendar_or_meeting (Zoom reminder, appointment reminder) | 3 |
fyi (Notion weekly digest) | 1 |
Six of the replies were tagged high-priority (mom's dinner invitation, the designer's mockups, the legal contract, a PR review, a colleague asking to reschedule a roadmap sync). Six medium. Thirteen low.
What the drafts actually look like
The point isn't just categorization — it's that the agent wrote usable drafts. Not generic "Thanks, I'll get back to you" filler. Actual replies that sound like a person who read the email.
To mom, about Sunday dinner:
Hi Mom — I'd love to come Sunday at 6! The lamb dish sounds amazing, and I'm excited to hear about Dad's new project. See you then, xo
To the designer following up on mockups:
Hey Sarah — v2 looks great, thanks for tightening the nav bar. I can do 15 minutes tomorrow morning around 10am if that works. If not, I'll leave async comments on the pricing table by end of day.
To the lawyer about the redlined MSA:
Thanks Ada — the redlines look good and the 30-day termination window is what I asked for. On section 7.3, let's talk it through before signing; could we do a quick call tomorrow?
To the security alert about a new device signing into my password manager:
Thanks for the alert — that was me signing in from Chrome on my Mac in San Francisco. No further action needed.
That last one is the tell. The agent noticed the alert was about a MacBook sign-in from San Francisco, matched it to context that made sense, and closed the loop — instead of drafting some vague "let me check this" reply.
The receipt
| Emails processed | 25 |
| Drafts generated | 11 |
| Wall-clock time | 4 minutes 9 seconds |
| Actual cost via Reality Router | $0.012 |
| Same workload on Claude Sonnet 5 | $0.31 (26× more) |
| Same workload on Claude Opus 5 | $0.52 (43× more) |
| Same workload on Claude Fable 5 (top-tier) | $1.04 (86× more) |
Cent-and-a-quarter for 25 emails triaged and 11 replies drafted. If you'd sent the exact same tokens to any of the flagship models, you'd have paid between a quarter and a dollar for the same output. That's the router's job — picking the right-sized model for each of the seven agent calls this took, so a 40x-plus price gap disappears without any thinking on your end.
Why triage, not "reply-and-send"
The agent could send these replies for me. Some tools do. I've deliberately built the workflow to stop one step short: draft ready, waiting for me to hit send.
The reason: replies are relationships. The mom draft is warm because that's how mom is talked to; the investor draft is professional-but-not-eager because that's the right frame there; the lawyer draft flags section 7.3 for a real conversation instead of just accepting. Those are calibrations only I can make. But once I set the tone in five one-word tweaks per draft, the mechanical part — remembering context, writing the polite opening, matching the register of the incoming email — is done.
Twenty-five emails, twenty minutes of my attention saved, eleven drafts ready to review in about the time it takes to make coffee.
What the router did
Seven agent calls total. Reality Router split them between DeepSeek's fast cloud model (five calls, for the bulk categorization + drafting) and gpt-4o-mini (two calls, for the reasoning-heavier decisions). No expensive-model routing needed for a task this shape.
That's exactly the routing behavior you want for a workflow you'd run every morning: fast, cheap, correct model per call, no thinking required from you about which tier to use.
Try it yourself
Reality Router is open source. Point OpenCode (or any OpenAI-compatible tool) at it, plug in your own API keys, and hand it your inbox instead of reading it top to bottom yourself.
Test run used a 25-email synthetic inbox with realistic subjects, senders, and bodies. If you're pointing the same workflow at your real inbox, connect via IMAP or your email provider's API; nothing in the setup requires the emails to leave your machine or pass through a third party besides the LLM API of your choice.