|
// the signal · issue 002
the agent you couldn't afford last month just got cheap.
Last month, running an AI agent on your busywork cost too much to bother.
This week it costs about a third of that. Same work. The math flipped.
| > the_signal CLAUDE SONNET 5 DROPPED |
|
Anthropic shipped Claude Sonnet 5 on 30 June. Every headline talked benchmarks. Wrong lens.
The one that touches your bank account: intro pricing is $2 in / $10 out per million tokens, about a third of what the top model costs, for performance Anthropic itself puts close to Opus 4.8. That window shuts 31 Aug, then it steps to $3 / $15.
why it matters for your money: the intelligence didn't jump, the unit cost dropped. Inbox triage, lead sorting, first-draft quotes, research: the agents that were too expensive to run all day were priced for a lab. Now they're priced for a business.
how to use it this week: a cheaper model you never run is just a smaller number on a page. Move a real task onto it, or the price cut means nothing.
|
| > the_play MOVE ONE TASK |
|
Don't go explore AI. Pick the single most repetitive thing a person gets paid to redo every day. The boring one holds the cash. The three that pay back fastest:
1. inbound lead sorting
Every enquiry scored and routed, so the good ones get answered in minutes.
2. first-draft proposals and quotes
The 80% skeleton written before you touch it.
3. customer-email triage
Sorted, tagged, draft replies waiting on your yes or no.
Pick ONE. Run it on Sonnet 5 this week, measure hours a human took versus hours now, and lock the intro rate before 31 Aug. A 48-hour test, not a six-week build.
|
| > the_steal THE CHEAP SWAP |
|
Re-run an automation you already have on Sonnet 5.
Same inputs, same prompt. Put the output and the bill side by side. Most of the time you keep the quality at roughly a third of the cost, and you find that out today instead of next quarter. One swap, wider margin on everything you've built.
|
| > also_this_week 30 SEC, YOU'RE CAUGHT UP |
|
→ The frontier models Washington locked down are back. Anthropic restored Fable 5 globally on 1 Jul and re-opened Mythos 5 for select US orgs. Access flipped on a policy call nobody voted on. Keep a fallback wired in.
→ OpenAI floated a 5% stake to the US government, and GPT-5.6 is still on hold. Don't roadmap around a model that isn't shipped yet.
→ Microsoft stood up a new $2.5B arm, Frontier Company, to make enterprise AI deployments work. Even Microsoft thinks the hard part is deployment, not the model. That gap is your opening.
→ The market flipped from "tokenmaxxing" to efficiency. The winners aren't spending the most on AI. They're getting the most output per pound. Sonnet 5 just made that easier.
|
The tools got cheap this week. Knowing that is worth nothing. Moving a task onto them before the price runs out is the whole edge.
Hit reply: what's the one job you'd hand a robot first? I read every reply.
Talk soon,
Mike
|