Good morning.

Salesforce hired zero new software engineers in its last fiscal year. The freeze ran across the entire company, and revenue kept climbing the whole time.

Same window, Wix cut 20% of its workforce, GitLab cut 14% and pulled out of 22 countries, and hiring across the US labor market dropped to its lowest level since 2010. The mass-layoff headline everyone keeps waiting for never came but the hiring freeze did.

The operators capturing the upside are running one move: they grow output without growing payroll, and they can do it because the cost of the work itself is collapsing underneath them. The price of intelligence is falling faster than almost anyone is pricing into their plans.

Here are 15 important updates from the last few weeks, each with my read on what it means for an operator at your scale.

— Sam

Let’s get into it.

Marc Benioff confirmed Salesforce added no new software engineers in fiscal 2026, expanding engineering capacity through AI instead. He cited a 30% productivity lift among existing developers using Agentforce and Claude. The company also trimmed its support team from 9,000 to 5,000, moving the rest of the load to autonomous agents, and committed $300 million in Anthropic token spend. (Source)

Salesforce is still hiring salespeople. Engineering and support are the tracks they froze. For an operator your size, that's the move worth copying, and you don't need layoffs to run it.

Freeze the next requisition, let attrition do the trimming, and route the new volume to agents. The question to ask before every hire from here: what would it take for an agent to absorb this instead? Your next $1M in revenue probably doesn't require the headcount your last $1M did.

Macro data shows AI's main effect on the workforce so far is a structural hiring freeze rather than mass layoffs. Tech layoffs passed 134,000 in the first five months of 2026, but hiring reached its lowest level since 2010. The Federal Reserve's Beige Book reported firms limiting headcount through attrition while using AI to keep existing workers productive enough to avoid new hires. Entry-level software and customer service postings are down roughly 20%. (Source)

If you're waiting for a dramatic layoff headline to mark the moment AI changed the labor market, you'll miss it. The change lives in the job that never gets posted and the intern class that gets canceled.

For your business this cuts both ways. Established players scaling output without adding payroll raise the bar a new competitor has to clear, and they raise it for you too. The companies adding the least headcount per dollar of revenue are setting the pace now, and the gap compounds quarter over quarter.

Sonora, an online guitar school, cut its headcount from 48 to 30 without losing revenue, saving roughly $250,000 a year. Founder Spencer Handley used Claude to build lightweight internal software that replaced HubSpot, Calendly, Vimeo, and DocuSign. Sonora also automated its outbound setter team, customer onboarding, and marketing operations with custom agents. (Source)

This is the most useful SMB case of the year because the scale matches yours. Handley used AI to build a bespoke operating system for one specific business and deleted thousands in monthly seat licenses on the way out.

Your SaaS stack is a tax on margin that compounds every year you renew it. At $1M to $10M, that stack can stop being a pile of subscriptions and start being a database your own agents run on. Start with the one tool whose invoice you resent most, and see how much of it you can rebuild.

A leaked internal Meta dashboard called "Claudeonomics" ranked the company's 85,000-plus employees by individual token consumption. The top developer burned 281 billion tokens in 30 days, equal to hundreds of thousands of dollars in compute. Meta shut the dashboard down after the leak. (Source)

The high-output developer of 2026 manages compute more than they write code. One person running agent farms can ship the volume of a whole traditional engineering team in a day.

If you hire engineers on how fast they type, you're measuring the wrong thing. The skill worth paying for is the ability to orchestrate and budget a swarm of agents toward a result. The person getting three weeks of work done in an afternoon is doing it with leverage you can measure, hire for, and teach to the rest of your team.

Uber set a $1,500 monthly cap per employee on agentic coding tools like Claude Code and Cursor. The cap followed a disclosure from Uber's CTO that the company spent its entire annual AI budget in four months after telling engineers to use AI as much as possible on internal leaderboards. Uber's COO also questioned the returns, noting it's hard to draw a line from token spend to shipped features. (Source)

Unmetered AI access has a ceiling, and Uber found it fast. A developer running five parallel agents across five checkouts can burn $400 in an afternoon.

If you're scaling a technical team, put budget guardrails in before the bill arrives, not after. Productivity only counts as a win when the margin gain clears the compute cost, and most teams aren't watching that line yet. Give people real leverage and a real number they're spending against.

DeepSeek locked in its promotional 75% price cut on its flagship V4 Pro model, setting output pricing at $0.87 per million tokens. The company reports that's roughly 7x cheaper on inputs and 17x cheaper on outputs than Claude Sonnet or GPT-5.5-Med, with cached read costs up to 87x cheaper than Western clouds. (Source)

The price floor under intelligence keeps dropping, and it changes your math more than any single model release does. If you're running single-turn chatbots, this barely registers.

If you're deploying agents that loop for hours through your data and code, token use compounds, and so do the savings. You probably won't self-host, and you don't have to. The point for an operator your size is that the commodity layer of your operation is getting cheap fast, which frees budget for the layer only you own.

Flo Crivello, CEO of Lindy, said the company moved all of its production AI traffic from Anthropic's Claude to the open-weight DeepSeek V4 model. Crivello reported better performance on Lindy's core use cases and savings of millions a year in token costs. It's one of the first public defections of a high-volume agent startup from closed Western APIs to open weights. (Source)

A venture-backed company moving its whole stack to open weights to save millions tells you the migration question is now live at the top of the market. You're not Lindy, and you don't run their volume. The lesson still applies.

The model under your agents is becoming a swappable part, not a permanent commitment. Build so you can route work to whatever's cheapest and good enough, and reserve the expensive frontier model for the small share of tasks that truly need it. Lock-in is a choice now, and an expensive one.

Pinterest CTO Matt Madrigal said the company moved its primary AI workloads off closed models to an open-source strategy. Pinterest post-trained Alibaba's open-weight Qwen model on its proprietary "taste graph" of user interests. The customized model delivers frontier-level recommendation quality at a 90% reduction in infrastructure cost versus closed APIs. (Source)

This is the playbook for any business sitting on data. Pinterest took a cheap, capable base model and trained it on their most valuable asset: the interest data only they have.

You likely won't fine-tune a model this year, and the principle still holds. It's the Data-as-Asset thesis in one example. The years of customer interactions, purchase histories, and support logs you've been treating as exhaust are the thing that gets more valuable as models get cheaper. Own that, and the price of the model stops being your problem.

Legal AI platform Harvey and Fireworks AI tested a hybrid setup that pairs an open-weight worker model (GLM 5.1) with Claude Opus 4.7 as a callable advisor, reserving the frontier model for the hard calls. On Harvey's Legal Agent Benchmark, the hybrid passed 18 of 100 tasks at full rubric for $368, against Claude Opus 4.7's 14 of 100 at $954. That's a 29% quality gain at 61% lower cost. (Source)

The winning architecture is a router. Easy tasks go to cheap models, and the heavy reasoning models get saved for the top slice that needs them.

If your developers pipe every prompt to a premium frontier model out of habit, you're paying a margin tax for no extra quality. A simple routing layer can cut infrastructure cost by half and improve results at the same time, because the expensive model stops getting handed work a cheap one does fine. Ask your technical lead where every prompt currently goes. The answer is usually "the most expensive option, always."

Wix CEO Avishai Abrahami said the website-building platform is cutting roughly 20% of its workforce, just over 1,000 people. He named AI's rapid improvement as the primary driver, letting smaller teams handle work that used to need large engineering and support cohorts. He predicted 70% of the top 20 most common US jobs will be altered or displaced by AI within five to ten years. (Source)

Wix rebuilt its operating model around AI in a single quarter, and it's a mature company rather than a startup. You're competing against businesses shedding a fifth of their headcount while holding or growing output.

You don't have to match their layoffs to keep pace. The part to copy is the redesign: make sure your next stage of growth doesn't require the hiring your last stage did. The companies doing this aren't smarter than you. They started reworking the actual work earlier, and that head start closes the day you start too.

GitLab said it's laying off about 350 employees, 14% of its full-time staff, and pulling operations out of 22 countries, a 37% cut to its geographic footprint. The company is reorganizing to serve AI workloads, even after reporting a 23% jump in first-quarter revenue. (Source)

Profitable, growing companies are restructuring too, so read this as something other than cutting losses. The geographic exit is the part to sit with.

Maintaining legal entities and local payroll in 22 countries is operational drag once agents can handle localized support and compliance. Look at your own footprint the same way. The overhead you carry out of habit is worth a fresh audit when a smaller, more centralized operation can cover the same ground at a fraction of the cost.

Administrative AI startup Lassie raised a $35 million Series A led by Andreessen Horowitz, bringing total funding to $47 million. Its agents run back-office paperwork end to end, starting with dental practices: pulling insurance reimbursements, reconciling them against records, and updating the system of record rather than assisting a human through it. Lassie now runs in more than 700 practices across 49 states and reports saving owners over 250,000 hours of admin a year. (Source)

The category moving fastest now is the autopilot. A copilot waits in the sidebar for you to type. An autopilot takes the login and runs the whole process while you sleep.

If you operate in an admin-heavy niche, stop shopping for tools that make your admin staff faster and start looking for systems that run the administrative role outright, freeing your people for the client relationships a machine can't hold. Lassie going straight into the insurance portals to pull and reconcile payments shows what running the role outright looks like, and that's a headcount line on your P&L, not a software subscription.

A Chinese startup launched Monako Glass, billed as the first wearable Linux computer built into a pair of glasses. The 48-gram device runs a custom Linux OS and executes terminal commands, Claude Code, and Codex on a heads-up display, with voice and gesture control. Unlike consumer smart glasses that act as notification screens, it's designed as a working developer station worn on the face. (Source)

While the big platforms chase a consumer metaverse, the practical use of AR is starting with developers. Watch where the builders go first, because that's usually where the real tooling shows up a year before everyone else notices.

If your team can monitor systems or debug while walking around, the meaning of "office hours" and "workspace" loosens. This one is early, and easy to file under novelty. File it under direction of travel instead.

Tesla launched its Robotaxi service in Austin with no safety monitor in the vehicle, the first time it has put riders in driverless cars with no Tesla staffer aboard. The rollout mixes a small number of unsupervised vehicles into the broader monitored fleet, with the ratio set to grow over time. (Source)

The physical world is catching the same cost curve the digital world has been riding. For e-commerce brands and logistics-heavy operators, driverless transport stopped being a five-years-out slide and went live on real streets.

The cost of moving goods and people is heading toward the same collapse we've watched in tokens. If your model leans on the current price of physical movement, start sketching what it looks like when that price falls hard, because Austin says the timeline is shorter than your last forecast assumed.

In an OpenAI research challenge, an autonomous agent named Aiden ran for 22 days and outperformed all 1,016 human researchers who entered. OpenAI couldn't extend a job offer to a piece of software, but the result moved through the technical recruiting world as a marker of how fast autonomous research is closing on top human work. (Source)

When the best candidate to build the next model is a model, human engineering becomes the bottleneck in its own pipeline. For an operator, the read is sharp.

If your core value is basic research, data synthesis, or standard software work, an agent will do it faster and cheaper soon, and your customers will figure that out.

Move your moat to the things agents can't copy: the proprietary data you've accumulated, the physical assets you control, and the trust you've built with customers over years. That's the asset that compounds while the cost of raw capability keeps falling.

The cost of raw capability is collapsing, and the companies winning are the ones turning that into output without turning it into payroll.

They freeze the hire, route the work, and put the savings into the part of the business a model can't reproduce.

That last part is the whole game. When intelligence is cheap and getting cheaper, your edge is the data only you own and the judgment only you can encode.

Until next time,
Sam Woods
The Editor

.