U.S. employers added 29,000 jobs in September and the unemployment rate was 4.2 percent, the Bureau of Labor Statistics (BLS) reported Friday ^[1]. BLS said both measures changed little in the month ^[1].

The September gain followed an average monthly gain of 45,000 over the prior 12 months ^[1]. BLS revised July down by 31,000, from +21,000 to -10,000, and August down by 29,000, from +162,000 to +133,000 ^[1]. The unemployment rate has stayed between 4.1 percent and 4.3 percent since March, and 7.1 million people were unemployed in September ^[1].

Average hourly earnings for private-sector employees rose 5 cents, or 0.1 percent, to $37.81, and are up 3.0 percent over the past 12 months ^[1]. Health care added 17,000 jobs, slower than its 12-month average gain of 33,000, and construction added 11,000 ^[1]. BLS described construction as little changed, along with retail trade, leisure and hospitality, and manufacturing (+9,000) ^[1]. Financial activities was little changed at -7,000, and the average private-sector workweek stayed at 34.4 hours ^[1].

On September 16, the Federal Reserve’s policy committee voted 12 to 0 to raise its target range for the federal funds rate, the short-term benchmark interest rate, by 1/4 percentage point to 3-3/4 to 4 percent ^[2]. Its statement said job gains have kept pace with the workforce and that the increase will support a timelier return to the committee’s 2 percent inflation goal ^[2].

By the numbers

• +29,000 Change in total nonfarm payroll employment, September 2026
• 4.2 percent U.S. unemployment rate, September 2026
• $37.81 Average hourly earnings, all private nonfarm employees
• 3.0 percent Increase in average hourly earnings over the past 12 months

Why it matters for your business (analysis)

Analysis. For an owner who hires, an unemployment rate that has held between 4.1 and 4.3 percent since March describes a labor market that has not moved sharply in either direction. Pay growth of 3.0 percent over 12 months is the average increase across private employers, a reference point for what other employers are paying compared with a year ago.

The downward revisions mean the earlier readings for July and August overstated job growth, and July now shows a loss of 10,000 jobs ^[1]. A flat 34.4-hour average workweek indicates employers did not add hours either ^[1]. Lenders often tie variable-rate business loans and credit lines to short-term benchmark rates, so a Fed increase can raise borrowing costs for a business with that kind of debt. Where rates go from here depends on data such as this report and the next inflation reading.

What to watch

• October 14, 2026: BLS releases the Consumer Price Index, the main measure of consumer inflation, for September ^[3].
• October 27 to 28, 2026: the Federal Reserve’s next policy meeting ^[4].
• November 6, 2026: BLS releases the October jobs report ^[5].

Sources

1. The Employment Situation: September 2026U.S. Bureau of Labor Statistics, Oct 2, 2026Primary
2. Federal Reserve issues FOMC statementBoard of Governors of the Federal Reserve System, Sep 16, 2026Primary
3. Release calendar, October 2026U.S. Bureau of Labor Statistics, Oct 2, 2026Primary
4. FOMC meeting calendars and informationBoard of Governors of the Federal Reserve System, Oct 2, 2026Primary
5. Employment Situation release scheduleU.S. Bureau of Labor Statistics, Oct 2, 2026Primary

The Internal Revenue Service and the Treasury Department published temporary regulations on Trump accounts in the Federal Register on September 30, 2026, effective that day ^[1]. A Trump account is the law’s name for an individual retirement account (IRA) with some special rules ^[1]. The regulations cover general requirements for the accounts, how an initial account is established, including automatic enrollment by the Treasury Secretary, and a type of contribution called a qualified general contribution ^[1].

The accounts come from Public Law 119-21, known as the One, Big, Beautiful Bill Act ^[1]. The document says contributions could first be made on July 4, 2026 ^[1]. It also describes a pilot program under which Treasury will pay $1,000 into the accounts of eligible children, a group that includes U.S. citizens born in 2025 through 2028 who meet the listed conditions ^[1].

For employers, the law added section 128 of the tax code, which says employers may contribute to the Trump account of an employee or an employee’s dependent ^[1]. Those contributions are excluded from the employee’s income up to $2,500 a year, adjusted for inflation for taxable years after 2027, and they count toward a $5,000 annual contribution limit ^[1].

The employer rules are in a separate proposal. The IRS published proposed regulations on August 11, 2026 on employer contributions to Trump accounts, including nondiscrimination rules, which limit how a benefit can favor certain employees ^[1]. The September 30 document says future guidance will address other issues ^[1].

By the numbers

• $2,500 Annual limit on employer contributions excluded from an employee’s income, adjusted for inflation after 2027
• $5,000 Annual contribution limit that employer contributions count toward
• $1,000 Pilot program payment Treasury will make to eligible children’s accounts
• July 4, 2026 Date the document says contributions could first be made

Why it matters for your business (analysis)

Analysis. The employer piece is optional. The law says employers may contribute, and no source describes a requirement to do so ^[1]. A business that wants to offer the benefit would be working from proposed rules, because the August 11 proposal has not been finalized in the September 30 document.

The document’s summary says the temporary rules affect trustees, account beneficiaries, responsible parties and donors who fund qualified general contributions. Employers are not in that list ^[1]. The nondiscrimination tests in the August proposal bear on how an employer could structure a contribution, and those tests are not yet final.

What to watch

• Final regulations on employer contributions under section 128. The September 30 document gives no date for them ^[1].
• Further guidance on other parts of the Trump account rules, which the document says is coming without a date ^[1].
• Rules on what Trump account money can be invested in. The IRS published a separate proposal on that topic on August 21, 2026 ^[1].

Sources

1. Trump Accounts (TD 10056), temporary regulationsFederal Register, Internal Revenue Service and Department of the Treasury, Sep 30, 2026Primary

I tried to trace them to a study. I could not. Here is what I found, and here is what you can quote instead.

The numbers to retire

I could not trace any of these to a study that supports the claim.

• 62%calls unanswered
• 85%never call back
• 80%hang up, no voicemail
• 78%buy from the first responder

• “62% of calls to small businesses go unanswered.” The closest thing to an origin is a 411 Locals sample of 85 businesses that counted voicemail as unanswered. Eighty-five shops is not an industry, and a voicemail count says nothing about what the caller did next.
• “85% of callers who can’t reach you won’t call back.” No study I could find.
• “80% of callers hang up without leaving a voicemail.” Same story.
• “78% of customers buy from the first business to respond.” Its cousin is “35 to 50% of sales go to the first responder.” They disagree with each other, and neither has a study behind it.
• Any “dollars lost per missed call” figure. Every one I found was a modeled example, somebody’s assumptions multiplied together. That is arithmetic, not a measurement.
• “Calls convert 10 to 15 times better than forms.” It cites BIA/Kelsey, 2015, a source that cannot be opened, so nobody can check it.

A fake number does more than fail to help. It sets the wrong expectation. If you believe 85% of missed callers are gone for good, you will plan around a problem you have not measured.

The numbers that hold up

The real data is thinner than the zombies. Most of it comes from software companies analyzing their own customers. I label it vendor data every time, because it is. It is useful and it is not neutral, and it skews toward shops that already run software. I am not recommending any of these companies. They are just the ones who published numbers.

• 42%of calls booked by the typical trade shop¹
• 52%of home-services callers spoke with a person²
• 14%of home-services leads were missed³

Booking rate: ServiceTitan, 2022 (vendor data)

In October 2022, ServiceTitan published call booking data from more than 3,000 US and Canadian trade businesses on its platform, covering calls from June 2022. The typical shop booked 42% of its calls. Shops with fewer than 5 techs booked 24%. Shops with 25 or more techs booked 59%. The data does not say why the gap is that wide. Size, staffing, and who picks up the phone all travel together.^[1]

Reaching a person, then asking: Invoca, 2026 (vendor data)

In July 2026, Invoca’s Home Services Lead Conversion Benchmarks Report covered more than 70 million calls across 10 industries, all from Invoca customers. Of callers to home services businesses, 52% spoke with a person. By sub-industry that ranged from 32% to 74%. And 55% of home services businesses don’t ask leads to book. Answering is not the same as selling.^[2]

Missed leads: CallRail, 2025 (vendor data)

CallRail’s 2025 benchmark report studied 1.1 million de-identified conversations across 7 industries and found that 14% of home-services leads were missed. Compare that with the 62% you have been quoted. Do not stack it against Invoca’s 52% either. They count different things on different customers. But 14% is a number somebody actually counted.^[3]

Speed: Harvard Business Review, 2011

In 2011, Oldroyd, McElheran and Elkington audited 2,241 US firms for Harvard Business Review (this link goes to a CustomerThink page covering the study). Only 37% replied to the lead within an hour. 23% never replied at all. The average reply took 42 hours. Firms that answered within an hour were nearly 7 times as likely to qualify the lead as firms that waited one more hour, and more than 60 times as likely as firms that waited 24 hours or longer. This was an audit of online inquiries across many kinds of companies, not phone calls to trades, so read it as proof that speed matters, not as a number for your shop.^[4]

What to measure in your own shop

Quoted numbers cannot tell you what your phone does. Your call log can. Pull one month from your phone carrier or phone app and count three things for calls from people asking for work:

1. Calls in. Every inbound call from a new or potential customer. Skip vendors, suppliers, and wrong numbers.
2. Calls answered. Calls where a person talked to the caller. Voicemail does not count.
3. Calls booked. Calls that ended with an appointment or a scheduled estimate.

Divide answered by in, and booked by in. Those are your answer rate and your booking rate. If you can, split them into before and after 6 p.m.

Then hold yourself up against the benchmarks above, with a grain of salt. ServiceTitan’s typical shop booked 42%, but its counting rules may not match yours. Beating a benchmark matters less than knowing your own number. Nobody has measured it for you.

The best evidence: Harvard Business Review, 2011

In 2011, Oldroyd, McElheran and Elkington audited how 2,241 US firms handled inbound leads for Harvard Business Review (this link goes to a CustomerThink page covering the study). What they found:^[1]

• 42hoursaverage time to reply¹
• 37%of firms replied within an hour¹
• 23%never replied at all¹

• Firms that replied within an hour were nearly 7 times as likely to qualify the lead as firms that waited one more hour.
• Against firms that waited 24 hours or longer, the multiple was more than 60.

Two caveats. It audited online inquiries across many kinds of companies, not phone calls to trades. And it compares groups of firms, so other differences between fast and slow companies could be mixed in. Still, it is the largest audit I found, at 2,241 firms. Waiting even one more hour cost a lot, and waiting a day cost far more.

The famous numbers are weak

You have probably seen “respond in five minutes.” Here is where that comes from:

• The “MIT” study, 2007. Commonly labeled the MIT study, this lead response study is vendor data tied to InsideSales. It reported that contacting a lead within 5 minutes instead of 30 made contact 100 times more likely and qualification 21 times more likely. It covered six companies. Six.^[2]
• Velocify, 2012. A vendor analysis of about 3.5 million leads. It said a call in the first minute lifted conversion by “almost 400%” and that 93% of converted leads were reached by the sixth call. This is vendor data from 2012, and I have no working link to the original, so I cannot check it.^[3]

Weak does not mean wrong. It means you should not build a shop rule on “100 times” and repeat it to your crew like scripture.

So how fast?

This part is my judgment, not data.

• During business hours, aim for minutes. Call. If they do not pick up, send a text right away so they know you saw the request.
• Treat one hour as the ceiling. The one solid study says the damage starts there.
• After hours, do not let it sit until noon. A quick acknowledgment that night beats a perfect call the next afternoon.

Time your own shop

Most owners have never timed this. Pick one week. For every new lead, whether a call, form, text, or message, write down when it came in and when a person first answered. Average the gaps. That is your number. The average in the HBR audit was 42 hours. If you cannot say “under an hour” with a straight face, you just found something worth fixing.

The benchmark: ServiceTitan, 2022 (vendor data)

In October 2022, ServiceTitan published call booking rates from more than 3,000 US and Canadian trade businesses on its platform, covering calls from June 2022. That is vendor data from its own customers, so it leans toward shops that already run software. It is also the largest trades-specific set I found.^[1]

• 24%booked by shops with fewer than 5 techs¹
• 42%booked by the typical shop¹
• 59%booked by shops with 25 or more techs¹

Two things jump out. Trade barely matters: plumbing, electrical, and HVAC land within five points of each other. Size matters a lot: the smallest shops booked less than half of what the biggest ones did. The data does not say why. My guess is that big shops have somebody whose job is answering the phone, and small shops have a person under a sink. That is a guess.

After 6 p.m. the number falls off

• 21%booked after 6 p.m. by larger shops¹
• 9%booked after 6 p.m. by smaller shops¹

In the same ServiceTitan data, booking fell to 21% for larger shops and 9% for smaller ones after 6 p.m. How many calls is that? A separate ServiceTitan look at residential HVAC in 2025 found 14.1% of inbound calls came after hours in June and 9.8% in October. ServiceTitan’s definitions of “after hours” and “after 6 p.m.” may not match, so do not multiply them together. The takeaway is plain enough: after-hours calls are a minority of volume, at least in HVAC, and the weakest part of the booking rate.^[2]

How to compute your own

1. Pick a window. Four weeks or more, so one strange week does not fool you.
2. Count calls in. Every inbound call from someone asking for work. Skip vendors and wrong numbers.
3. Count calls answered. Calls where a person talked to the caller. Voicemail does not count.
4. Count calls booked. Calls that ended with a scheduled appointment or estimate.
5. Divide booked by calls in. That is your booking rate.
6. Split by time of day. Before 6 p.m. and after. Your after-hours number will probably hurt.

Example math, not a statistic: 50 calls in, 30 answered, 17 booked. Booking rate is 17 divided by 50, or 34%. Answer rate is 30 divided by 50, or 60%. Those two numbers point at different problems. Calls you never answered are a coverage problem. Calls you answered and did not book are a conversation problem. Invoca’s 2026 report on more than 70 million calls (vendor data) found that 55% of home services businesses don’t ask leads to book. Ask.^[3]

What counts as good

Use the averages above as rough yardsticks, not grades. They are averages from one vendor’s customers.

• Around 24% with a small crew: you match the average small shop, and there is room to improve.
• Around 40%: you match the typical shop of any size, and sit well above the average for crews under 5 techs.
• 55% or more: you are close to the 59% ServiceTitan saw at shops with 25 or more techs. That is hard to do, and worth studying if you get there.

Your own number, tracked month over month, tells you more than any benchmark.

The U.S. Census Bureau reported on October 1, 2026 that private residential construction spending was at a seasonally adjusted annual rate of $882.3 billion in August, 1.1 percent above the revised July estimate of $872.7 billion ^[1]. Census lists the change as 1.1 percent (±1.3 percent) and notes that when the range includes zero, it is uncertain whether spending rose or fell ^[1].

Compared with August 2025, private residential spending was down 4.8 percent ^[1]. Within it, new single-family construction was up 0.2 percent from July and down 3.5 percent from August 2025 ^[2]. The residential category also includes private residential improvements, which exclude rental, vacant or seasonal properties ^[2].

Total construction spending, which also counts public projects and commercial buildings, was $2,203.1 billion at an annual rate, 0.9 percent above July and 1.7 percent below August 2025 ^[1]. During the first eight months of 2026, spending amounted to $1,450.4 billion, 3.1 percent below the same period in 2025 ^[1].

Private nonresidential construction was at $773.0 billion, 1.0 percent above July ^[1]. Spending on data centers, a category of private office construction, was up 73.2 percent from August 2025 ^[2]. Census says the figures are adjusted for season but not for price changes, and that August numbers are preliminary and subject to revision ^[1].

By the numbers

• $882.3 billion Private residential construction spending, August 2026, seasonally adjusted annual rate
• 1.1 percent (±1.3 percent) Change from July, with Census’s margin of error
• -4.8 Percent change in private residential spending, August 2026 from August 2025
• -3.5 Percent change in new single-family construction, August 2026 from August 2025

Why it matters for your business (analysis)

Analysis. Trades differ in how much they depend on new building. Companies that install heating, plumbing, electrical or roofing systems in new houses follow construction activity closely. Companies that repair and replace systems in existing homes follow a different mix of demand. The new single-family figure speaks most directly to the first group.

The release does not separate remodeling from new-home building inside the residential total, so it cannot show how much of the 4.8 percent year-over-year drop came from each. Because the data are not adjusted for price changes, some of the dollar movement can reflect higher costs rather than more work ^[1]. The monthly rise is within the margin of error, so the year-over-year decline is the firmer signal in this release.

What to watch

• November 2, 2026: Census releases September construction spending ^[1].
• Revisions to August. Census says the average absolute change from preliminary to first revision for private construction is 0.75 percent ^[1].
• Census says it can take as long as 8 months to establish an underlying trend for specific categories of construction, so single months can mislead ^[1].

Sources

1. Monthly Construction Spending, August 2026 (CB26-158)U.S. Census Bureau, Oct 1, 2026Primary
2. Value of Private Construction Put in Place, seasonally adjusted annual rateU.S. Census Bureau, Oct 1, 2026Primary

OpenAI introduced GPT-6 Sol and GPT-6 Luna, two new AI models, on September 22, 2026, according to trade coverage ^[2][3]. OpenAI’s announcement page shows no publication date, only an update note dated September 29, 2026 ^[1]. OpenAI says it reduced API prices for both models by 50% compared with their GPT-5.6 promotional pricing ^[1]. The API is how software makers buy access to the models, billed per token, a token being a small chunk of text about the size of a word fragment.

Per million input tokens, Sol dropped from $4 to $2 and Luna from $0.20 to $0.10, according to OpenAI ^[1]. Output prices fell from $20 to $10 for Sol and from $1.20 to $0.50 for Luna ^[1].

OpenAI also says Sol makes about half as many mistakes as its predecessor, and that at high effort Luna improves on its predecessor by 5.4 percentage points at 58% lower cost per task on AutomationBench, a test of business tasks across apps ^[1]. These results come from OpenAI and have not been independently verified by Midwatch Radar. The models are available in ChatGPT Work and Codex on paid plans ^[3].

By the numbers

• 50% Price cut on GPT-6 Sol and Luna versus GPT-5.6 promotional pricing, per OpenAI
• $0.10 GPT-6 Luna input price per million tokens, down from $0.20
• $2 GPT-6 Sol input price per million tokens, down from $4
• 58% Lower cost per task for Luna at high effort on OpenAI’s AutomationBench

Why it matters for your business (analysis)

Analysis. Many business tools that small companies already pay for now include AI features, such as drafting replies, summarizing calls, tagging records or answering the phone. The software maker typically pays an AI provider by usage for each of those actions, so a lower per-token price lowers that cost.

Whether the saving reaches a subscriber depends on how each software company prices its own plans, and the sources do not cover that. The price cut applies to vendors that use OpenAI models, and other AI providers set their own prices. OpenAI measures the 50% cut against GPT-5.6 promotional pricing, so the comparison point matters when reading the figure ^[1]. The accuracy gains are OpenAI’s own claims.

What to watch

• Whether software vendors that build on OpenAI models announce price or plan changes. The sources list none.
• Independent testing of OpenAI’s accuracy and cost claims. As of October 2, OpenAI’s page is the source for those figures ^[1].
• OpenAI’s page carries a September 29, 2026 update pointing to a newer model, GPT-6.1 Sol, so details on it may change ^[1].

Sources

1. Introducing GPT-6 Sol and LunaOpenAI, Sep 22, 2026PrimaryVendor data
2. OpenAI cuts GPT-6 prices in half with Sol and LunaThe Next Web, Sep 22, 2026Secondary
3. OpenAI adds GPT-6 Luna and Sol, touts lower pricesConstellation Research, Sep 22, 2026Secondary

Anthropic told developers on September 30, 2026 that Claude Sonnet 4.5, one of its AI models, will be retired on November 30, 2026 ^[1]. The company’s deprecation page lists claude-sonnet-5-5, or Claude Sonnet 5.5, as the recommended replacement for the Sonnet 4.5 model ID, claude-sonnet-4-5-20250929 ^[1].

Retirement means the model stops working. Anthropic’s documentation says requests to a retired model fail, so software still calling Sonnet 4.5 after November 30 would return errors instead of answers ^[1]. The page says Anthropic gives at least 60 days’ notice before retiring a publicly released model, and that it notifies affected customers by email and in the documentation ^[1].

The dates apply to the Claude API, which is the connection that software uses to send requests to Claude, and to Claude Platform on AWS and Microsoft Foundry ^[1]. Amazon Bedrock and Google Cloud are run by partners and set their own retirement schedules, so dates there can differ ^[1].

The same page lists Claude Sonnet 5.5 as active, with a tentative retirement date of not sooner than September 28, 2027 ^[1].

By the numbers

• November 30, 2026 Retirement date for Claude Sonnet 4.5 on the Claude API
• 60 days Minimum notice Anthropic says it gives before retiring a publicly released model
• September 28, 2027 Earliest tentative retirement date listed for the replacement, Claude Sonnet 5.5

Why it matters for your business (analysis)

Analysis. Most small businesses do not call an AI model directly. They use one inside a product from a software company, such as a website chatbot, a phone answering assistant or a tool that drafts replies to customers. If a vendor built that product on Sonnet 4.5, the switch to a newer model is the vendor’s to make before November 30 if the tool is to keep working.

Anthropic’s notice goes to developers, so a business owner hears about it only if the vendor mentions it. A swap can also change how replies read, which is why Anthropic’s page recommends testing replacement models well before the retirement date ^[1]. An owner can ask a vendor which AI model sits behind a tool and whether it is moving off Sonnet 4.5.

What to watch

• November 30, 2026: scheduled retirement of Claude Sonnet 4.5 on Anthropic’s own platforms ^[1].
• Amazon Bedrock and Google Cloud publish their own schedules for the same model, so vendors that run Claude through those services may have a different date ^[1].
• The page also lists Claude Haiku 4.5 with a tentative retirement date of not sooner than October 15, 2026, and Claude Opus 4.5 not sooner than November 24, 2026. Neither is marked deprecated ^[1].

Sources

1. Model deprecationsAnthropic (Claude Platform Docs), Sep 30, 2026PrimaryVendor data