The cheapest employee you will ever hire answers your phone. It costs less than a coffee a day.

The short version

• Google’s Gemini 4 Argon entered global availability on 14 October 2026 at $2 per million input tokens and $10 per million output, with a task mode that keeps working after you close the app.
• OpenAI’s Dots, from 12 October, ships always-on agents on their own cloud computers with persistent memory, to ChatGPT Pro and Business Premium subscribers first.
• Six hundred answered enquiries cost about $25.40 in tokens at full list price. Under a dollar a day.
• The work around the model still costs real money: wiring it to your systems, teaching it your rules and drawing the Authority Line. Token prices fell; that craft did not.
• So being early is worth something. The model reaches every competitor at the same price. The trusted version reaches only whoever does the work.

There are weeks in technology that quietly change what an ordinary business can afford. This was one of them. On 12 October 2026 OpenAI began rolling out Dots, always-on agents running on their own cloud computers, with persistent memory, that pursue goals between conversations. On 14 October, Google’s Gemini 4 Argon entered global availability with pricing that would have been laughable eighteen months ago: $2 per million input tokens, $10 per million output. Meta had already put Muse, its small business agent, in front of firms under 250 people.

Put those three together and the shift is clear: the big labs have stopped selling intelligence and started selling labour.

And almost no one has switched. Which means the door is open, and every month you leave it that way is a month your competitors handed you.

The week the door opened

Google’s task mode is the clearest signal. Give it an errand, book something, chase something, find something, and it keeps working after you walk away. It connects to 40-plus apps including Instacart, OpenTable, Adobe and Shopify. OpenAI’s Dots each get their own memory, so they gather context about your business the way a colleague does.

Then the price. At $2 per million input tokens, one talk, even a long one, with your records attached, costs a few cents. Several hundred of those lands in the tens of dollars.

The old questions are settled. It works. You can afford it. What is left is you.

What $2 actually buys

Here is the arithmetic, using published list prices. One enquiry handled end to end: the talk, the lookup against your records, the confirmation, the note afterwards. Generous assumptions, not flattering ones:

Twenty-five dollars and forty cents a month, at list price, assuming every enquiry is a long one. Call it under a dollar a day, less than the coffee you buy on the way in.

The expensive part of answering your phone is not intelligence any more. It is everything around it.

The phone nobody answers

The cost that hurts a small business is not on any invoice. It is the call that rang out, someone who wanted a price, a booking, an answer your website could not give them. When no one picked up, most did not ring back. They rang the next name on the list.

A missed call is not a voicemail waiting politely. It is somebody who needed a thing, decided you were busy, and gave their money to someone else.

Then the version that stings more, because this one was yours. You did answer, on a ladder, in a van, at a child’s party. You said “I’ll call you back” and meant it completely. Then the day happened, and the promise quietly rotted.

Someone carries it. It is not even their job, whoever is nearest the phone, whose own work now happens at seven at night. That is what a digital employee takes: not the judgment or the relationships, but the answering, logging and chasing, at eleven at night.

The part that is not free

The honest part now. The token is nearly free. Turning it into someone you would trust near your customers is still work, and none of it got cheaper:

• Wiring it to reality. The model does not know your inbox, your diary or who is on a payment plan. Someone connects it.
• Teaching it your rules. What counts as a valid refund? Which quotes need approval? That lives in your colleagues’ heads, not your documentation.
• Reviewing it honestly. Right 95% of the time is still a liability if the 5% is expensive.
• Drawing the line. Every workflow needs a boundary: below it the software acts, above it a person decides. Wrong either way and the money is wasted.
• Governing it. Access control, logging, and one switch that turns it off.

That is why most attempts have stalled. MIT’s NANDA report in August 2025 found roughly 95% of generative AI pilots never reaching measurable profit-and-loss impact, and blamed integration, data readiness and the workflow itself, not model quality, and never price.

Which is why being early pays. The model reaches every competitor at the same price. The trusted version reaches only whoever does the unglamorous work.

What the first movers get

So what changes in an ordinary business that switches this month?

• Sunday stops costing you Monday. What used to greet you at eight is already handled, acknowledged, answered, logged.
• No one falls through. No missed call, no broken promise to ring back. Every enquiry gets an answer, at whatever hour it came in.
• Your people get their jobs back. Whoever was nearest the phone stops being nearest the phone. The admin that ate their evening is no one’s job.
• Records get kept. Six months in, every chase had a date on it, and none of it depends on one person staying employed.

And the quiet one: you start winning the calls themselves. In most local markets, trades, clinics, hospitality, professional services, answering first is not an edge, it is most of the game. When a rival’s phone rings out on a Saturday and yours does not, you were there.

None of that needed a better model than the one that shipped on 14 October. It needed someone to decide, this month, to stop answering.

What the models still can’t do

The limits now, because optimism without them is sales. The Remote Labour Index, a 2026 benchmark of 240 real freelance projects from researchers at the Center for AI Safety and Scale AI, found the best agent completing 2.5% of projects end to end, with major models scoring between 0.8% and 2.1%. Given real briefs and deadlines, agents failed at the long tail: client communication, iterating on feedback, delivering work a person would accept.

Read it the useful way round. What they fail at is judgment work, the part with a client expecting taste, tact and someone accountable. That stays with people.

Which is the whole design brief. A digital employee should be allowed to do the work no one in your business wants, and forbidden, technically forbidden, not politely discouraged, from the work that needs a person behind it. The Authority Line is how. The model is ready. The discipline is the product.

Where Twinthos comes in

This is the change worth being cheerful about, and it has little to do with model names. Digital employees take the repetitive work, the answering, confirming, chasing, logging and filing, and do it day and night, without fatigue. People get back the parts only a person can do: reading a customer’s hesitation, deciding the exception, being kind on a bad day. Software takes the volume. Humans keep the judgment.

The operating layer is what Twinthos sells. Not the model, we do not charge for tokens and you never think about them, but the managed digital employee on top: wired into the email, diary and accounting systems you already run, taught your rules, working while you sleep, and stopping at the Authority Line where a person decides. Below it, it acts alone. Above it, it cannot.

Nothing to wait for. What shipped on 14 October can carry your phone today. What is missing is your decision, then a few weeks of someone doing the wiring properly. The £999 Operations Audit starts it: we measure where work stops unless someone remembers to push it, put the drag in staff-hours, and tell you plainly if it is not worth doing.

Next step

Start with the £999 Operations AuditBook a discovery call

Twinthos is a managed digital employee service, not software you install. It works across the email, diary and accounting systems your business already runs. Below the Authority Line it acts on its own; above the line a person decides.

Common questions

Is Gemini 4 Argon really $2 per million input tokens?

It was the introductory global pricing announced 14 October 2026, alongside $10 per million output tokens. Lab list prices change often; the figure is correct at the time of writing.

So what does answering the phone actually cost in a month?

In the worked example above, about $25.40 at full list price for 600 mixed enquiries, under a dollar a day. That is tokens alone; it excludes the wiring, teaching and review that make the software safe to leave alone.

If tokens are nearly free, why is a managed digital employee still expensive?

The expense is integration, configuration, review and governance, the staff work that makes a model useful in one specific business. Token cost is now a rounding error, and that craft does not get cheaper as models do.

What does the 2.5% Remote Labour Index result mean for businesses?

It measures agents finishing whole freelance projects end to end, including client-facing judgment work. The low score says they are not ready to replace client-facing roles, not that they cannot carry structured follow-through.

What happens if the digital employee gets something wrong?

Every workflow gets an Authority Line before it is switched on. Below the line the software acts alone; above it, a person decides. Nothing crosses without a human, and everything is logged with the reason why.

Sources

1. AI agent news round-up, AI Agent Store, week of 14 October 2026 Primary report of Gemini 4 Argon global availability, pricing, task mode and integrations; OpenAI Dots rollout and GPT-6 Astra details.
2. Meta introduces Muse for small business, Meta Newsroom, 29 September 2026 Primary source for Meta’s small business agent, its integrations and its free tier.
3. Meta launches Muse for small business, CNBC, 29 September 2026 Coverage of the Meta Enterprise Platform and the free-tier positioning for companies under 250 people.
4. Remote Labour Index coverage, BusinessDay, 2026 Report on the 240-project, 23-category benchmark by the Center for AI Safety and Scale AI; the 2.5% best-agent figure and the failure modes.