Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Pallet, a startup building shelters for homeless people, shut down — learn why below

• Why adding more AI agents makes your team slower

• GPT 6 Astra and Fable 5.1 are out now and they are impressive — learn what people have been saying and building below

A big thanks to today’s sponsor…

Your best decisions start with listening ^AD

Every decision shapes your company's trajectory. SurveyMonkey ensures you're making them based on real insight, not assumptions.

Gather customer feedback through natural conversation. Get instant AI-powered analysis. Benchmark your metrics against peers in your space.

No expensive research team needed. Just real data, real fast. Integrate with ChatGPT and Slack so feedback flows directly into your workflow.

The companies learning fastest from their customers grow fastest.

→ Learn how today

This Week In Startups

🔗 Resources

Why adding more AI agents makes your team slower

Nobody cares about your new product (sorry)

How to turn your AI into a world-class designer

Will ChatGPT recommend your brand? Run a free 2-minute AI readiness scan on your site. * 

📰 News

GPT has solved one of the Millenium Problems

Meta released a new personal AI agent called Muse

Apple announced a new foldable phone: the Iphone Duo

Nvidia confirms it will buy Hugging Face for $12.9 billion

💸 Fundraising

Gimlet Labs, an AI infrastructure startup building a multi-silicon inference cloud, raises ⁠$300M Series B funding

Aitan, a defense tech startup building autonomous robotic warfare systems, raises ⁠$41M funding round

Resect AI, an AI accountability platform for detecting hallucinations and controlling model behavior, raises ⁠$25M funding

Diameter Pay, a fintech infrastructure platform for global US dollar payments and virtual accounts, raises ⁠$10M Series A funding

* sponsored

Fail(St)ory

The Missing Village

Pallet built flat-pack sleeping cabins that cities could assemble in under an hour and group into staffed shelter villages.

On September 4, 2026, the company ceased operations after nine years. Pallet's own figures show annual village openings falling from 35 in 2021 to five in 2026.

What Was Pallet:

Pallet made private shelters for people experiencing homelessness or displacement. Its cabins had beds, windows, locking doors and optional heating or air conditioning. Kitchens and bathrooms were initially communal, which kept the units light enough to move and deploy quickly.

For residents, a lockable room offered privacy and security that a tent or congregate shelter could not always provide. For cities, it created shelter capacity faster than permanent housing and gave service providers a stable place to reach people.

Cities, counties, tribes and nonprofits bought the cabins, usually with public or philanthropic funds. A separate provider often ran the village, handling intake, meals, sanitation, security, case management, maintenance and housing placement. The cabin was one line in a much larger operating budget.

Pallet primarily generated project-based revenue from sales of shelters and village components. That made growth depend on large public purchases, with appropriations, procurement and site approvals sitting between interest and revenue.

Assembly could take less than an hour. However, opening a site could take two years once community opposition, zoning, utilities and operator readiness entered the process.

The gap showed up after customers had already spent money:

• Nashville bought 108 cabins for $1.2 million, then left most in storage for years while officials worked through code and operating requirements.

• Rochester's 19 cabins also sat unused as insurance, nonprofit capacity and funding stalled.

Pallet responded by refusing to ship without an essential-services plan and adding advisory work on funding, sites, policy and operators. Those moves were meant to reduce the risk of unused cabins, while pulling the company further into problems outside manufacturing.

During the pandemic, Pallet reported producing 1,222 shelters and opening 35 villages in 2021. It raised a $15 million Series A the following year and used outside capital for materials, factory space and hiring.

The later S2 cabin added safety, energy and comfort features. A furnished, delivered and installed unit cost roughly $16,000 to $17,000 in 2026.

Pallet still had buyers near the end:

• Seattle opened 50 units in June.

• Dallas selected Pallet for a proposed 75-unit site in August

• and existing villages remained open after the shutdown.

Those late projects show that demand remained through the end. Pallet never disclosed its order book, leaving its shrinking deployment count as the clearest measure of the wider decline.

The Numbers:

• 🏗️ Founded: 2016

• 🏠 Shelters shipped: 5,780

• 📍 Reach: 150+ sites in 139 communities

• 📉 Annual villages: 35 in 2021 to five in 2026

• 💰 Series A: $15M in 2022

• 🛑 Operations ceased: September 4, 2026

Reasons for Failure: 

• Public demand arrived in waves: Pallet expanded factory capacity and hiring near the pandemic-era peak. Village deployments then declined for several years. Seattle's mayor said shifting public funding left Pallet without enough purchasing demand, and California placed zero orders with Pallet after selecting it for a 1,200-home program. The likely consequence was a factory carrying fixed costs against a thinner project pipeline, although Pallet never disclosed its utilization or margins.

• Every sale depended on a second operating system: Land, utilities, code approval, insurance, services and permanent-housing exits determined whether purchased cabins became a working village. Delays left units idle and stretched the time from political commitment to repeatable revenue. Pallet's move into advisory services suggests these customer-side dependencies had become part of its sales problem.

• Product and mission choices raised the entry price: Early cabins cost $3,500 to $7,500, while later quotes reached roughly $16,000 to $20,000 for a more capable installed unit. Pallet also used paid, supported labor rather than volunteers. Those choices may have strengthened durability and social impact, but cash-constrained buyers could count doors more easily than lifecycle value.

Why It Matters: 

• Match growth to the buyer's clock: Before adding factory space, inventory or headcount, prove implementation-ready orders arrive fast enough to support the cash cycle.

• Treat external dependencies as part of the product: When customers need permits, infrastructure or partners before they can use it, solve those steps profitably or focus on buyers that already have them covered.

Trend

The AI Title Fight

Anthropic released Claude Fable 5.1 on September 1. OpenAI released GPT-6 Astra two days later. Since then, users have been conducting the industry's most orderly possible war: running the same prompt twice and arguing about the screenshots.

This is the closest thing frontier AI has to a main event. The models have roughly one-million-token context windows, the same 128,000-token maximum output, and identical list prices of $10 per million input tokens and $50 per million output tokens.

So, who is the winner?

Why it Matters

• The two releases are close enough that a leaderboard cannot make the buying decision for you. Artificial Analysis gives both models an intelligence score of 53, while Epoch scores Astra at 167 and Fable at 164, with heavily overlapping confidence intervals.

• Their identical token prices also produce very different bills. In Artificial Analysis's maximum-effort tests, Astra cost $3.26 per task and Fable cost $7.63. Fable generated about 78,000 output tokens per task, versus 27,000 for Astra.

What People Are Saying

• One builder had Astra play through Portal using screenshots for 23 hours and 43 minutes, making 3,336 tool calls. The bespoke setup says little about an ordinary coding task, but it gives Astra unusually concrete evidence of persistence.

• Simon Willison tested Fable by asking it to draw a pelican in SVG (a task AI usually struggles with), then increasing the model's reasoning effort. The quality rose sharply, but so did the bill. His max-effort result consumed 65,927 output tokens, took almost 14 minutes, and cost $3.30. Animating it cost another $1.37. It was, in fairness, a very nice pelican.

  

• Matt Shumer pushed both mdoels through Blender, Three.js, Unreal Engine, subagents, reference gathering, and visual critique. His launch-week builds included a 3D Manhattan, a browser neighborhood filled with zombies, and an

  agent civilization.

• Bart Slodyczka ran the clearest direct comparison: five prompts, five unattended app builds, and both models working from the same assignments.

• Tom Krcha used it to model a house. It seems that Astra is specially good at working with 3D spaces.

The Trend

So, what is the conclusion after a week? Which model seems to be better at which task?

• Astra looks like the better default when the model needs to act: operating software, using a terminal, navigating a browser, or building inside CAD and 3D tools. It leads the relevant independent benchmarks, finishes many public builds faster, and reached the same overall intelligence score with far fewer tokens.

• Fable looks stronger when the job depends on judgment: planning, reviewing, following an existing codebase, reasoning across long context, or adding visual polish. It may take longer and spend more, but builders often prefer what it notices and how the final result feels.

The evidence is less than ten days old and still too thin to turn these into permanent job descriptions. Codex and Claude Code shape the experience, public demos follow launch marketing, and some developers report the reverse.

Either way, Astra and Fable are currently the two models to beat. Artificial Analysis has them tied at the top, so the winner remains unclear. Everyone else, at least for now, is on the undercard.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico