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:

• Chimoney, a startup building a global payout API, shut down — learn why below

• LLMs are picking winners. Here’s how to become one

• There’s a new Claude model: Fable 5 — learn why you should try it below

A big thanks to today’s sponsor, SparrowDesk. Let their AI agents handle 60%+ of your support tickets automatically.

What if 60% of your support tickets never reached you? ^AD

As a startup, most of what lands in your inbox is repetitive. The same questions, the same responses, resolved manually by someone who should be closing deals or building your product.

SparrowDesk fixes this. Our AI agents auto-resolve 60%+ of queries before it reaches your team. One unified inbox for chat, email, and more, so nothing goes missed, and every customer gets a reply.

Early-stage teams get 90% off the first year through our startup program.

→ Try SparrowDesk for free

This Week In Startups

🔗 Resources

How AI changes the software value chain

Why you should focus on the Return on Tokens (ROT)

LLMs are picking winners. Here’s how to become one

Your startup didn't work out. Now what? * 

📰 News

Decart’s new world model can simulate hours of photorealistic driving

Google just fired a warning shot in the AI subscription price wars

OpenAI files confidentially for IPO, following Anthropic

Apple’s long-awaited AI Siri overhaul is finally here

💸 Fundraising

Arpio, an AI-native cloud recovery and resilience platform, raises $15M Series A

Forage, a financial infrastructure platform for government benefits payments, raises $40M Series B

Terra AI, an AI platform for mineral and energy exploration, raises $20M Series A

ZeroDrift, an AI compliance firewall for enterprise communications, raises $10M seed round

* sponsored

Fail(St)ory

Money Out, Not Money In

Chimoney tried to make messy global payouts feel simple.

It wanted to build a payout layer that could sit between companies, recipients, banks, wallets, local rails, and regulators across multiple markets.

It raised less than $1M to do it, which is roughly “nice seed round for a calendar app” money, not “move money across continents without breaking anything” money.

What Was Chimoney:

Chimoney was a payout orchestration layer for businesses sending money into fragmented markets.

It lived after the moment a company already owed someone money:

• A marketplace needed to pay sellers. 

• A remote-work company needed to pay contractors. 

• A creator or affiliate platform needed to pay users across countries where the preferred payout method changed by market.

That was the job Chimoney tried to handle. Integrate once, fund the payout, and let recipients receive money through the method that actually worked where they lived.

The product supported 41 currencies across North America, Africa, and Latin America. Customers could send bank transfers, mobile money, airtime, gift cards, and even stablecoins.

It was not trying to be a new Stripe. It did not help businesses accept card payments, run checkout, manage subscriptions, or process merchant payments. It focused on the other side of the flow: getting money out to recipients who might not have the same banking access, cash-out options, or local rails.

The messy part sat underneath that simple integration. If one recipient wanted mobile money, another needed a bank transfer, and another could only use airtime or a gift card, Chimoney had to make those options work behind the scenes. 

That meant routing payouts, managing local partners, handling failed transactions, reconciling balances, processing refunds, answering support tickets, and keeping enough liquidity in the right places. The customer saw one API. Chimoney carried the market-by-market mess underneath it.

Near the end, Chimoney tried to reposition part of this infrastructure around AI wallets and programmable payments. It was a cleaner story for the future, but the company still needed the payout business to work in the present.

By the time Chimoney stopped accepting new transactions in April 2026, the shape of the failure was clear: the product solved a real last-mile payout problem, but the company had to carry too much infrastructure before enough volume arrived.

The Numbers:

• 💸 Funding: Under $1M raised over the company’s life

• 🌍 Coverage: 41 currencies supported

• 🏢 Customers: Hundreds of businesses served

• 🛑 Shutdown: Stopped accepting new transactions on April 30, 2026

Reasons for Failure: 

• Distribution never became a real engine. The founder’s own explanation was direct: Chimoney had a technically solid, licensed platform, but no customer acquisition engine. In payout infrastructure, buyers need proof that money lands, failures get handled, compliance is covered, and switching risk is worth taking.

• The last mile was expensive to operate. Chimoney simplified payouts for customers by absorbing the operational mess itself. Every local rail, mobile money option, gift card route, off-ramp, refund path, and partner relationship added work behind the API.

• Fintech costs arrived before fintech scale. Under $1M is thin for a company moving money across multiple jurisdictions. Licenses, audits, banking partners, compliance, support, and liquidity management all cost money before transaction volume is large enough to carry them.

• Volume density never caught up. Payout infrastructure gets healthier when repeat transactions flow through the same rails often enough. Without strong distribution, Chimoney still had to maintain the platform, partners, support, refunds, and compliance systems on too small a revenue base.

Why It Matters: 

• Payout startups need dense, repeat transaction volume before the operational load becomes manageable.

• Developer experience helps, but payout buyers mostly care about whether money arrives and who fixes it when it does not.

• Future-market repositioning is hard when the current infrastructure business is already running out of room.

Trend

Claude Fable 5

There’s a new Claude model in town, so naturally, everyone is talking about it.

Anthropic says it’s SOTA, many developers claim it’s the new coding model to use, and a lot of people seem angry that it won’t help them make a bomb.

Here’s everything you should know about it: the good, the bad, and the biohazard disclaimers.

Why it Matters

• The best model is no longer always the model you get. Fable 5 is public. Mythos 5 is for vetted partners. Same family, different permissions, which means AI access is becoming a status layer.

• The coding hype seems earned. Fable 5 looks built for the annoying work developers actually hate: messy repos, migrations, broken tests, weird dependencies, and tasks that take more than one clean prompt.

• The guardrails are part of the product now. Sometimes Fable answers. Sometimes it refuses. Sometimes it quietly sends you to Opus 4.8.

The Coding Hype Has Substance

The strongest case for Fable 5 is coding.

Anthropic says it is better at long-running agentic work, large migrations, complex implementations, test writing, design fidelity, and reasoning across big codebases.

Early developer reactions mostly point in the same direction: it stays useful for longer, handles more context, and needs less babysitting once the task gets messy.

The specs help explain why. It reportedly has a 1 million token context window and 128,000 max output tokens. That gives it more room to read, plan, edit, test, and produce larger changes in one run.

Anthropic’s biggest example is Stripe using it to cut a 50-million-line Ruby migration from months to a day. My favorite example so far has to be this one:

The problem: it is very very expensive. Fable 5 costs twice as much as Opus 4.8, so letting it loose on a repo has the energy of giving a very smart intern your credit card.

The Guardrails Are Everywhere

The weirdest part of Fable 5 is that it is not just a model. It is a model with a bouncer.

Anthropic says Fable 5 uses the same underlying model as Mythos 5, but with stronger safeguards for public use. Mythos 5 is the more restricted version, available only to vetted partners, because it can help in areas that could be dangerous (mostly cybersecurity and biology).

So Fable 5 sometimes does something strange. If a cyber or bio prompt gets flagged, it may route the request to Opus 4.8 instead. 

The biology filters seem especially jumpy. Reports say Fable 5 refused or rerouted questions about mitochondria, mRNA vaccines, hay fever, asthma medication and antibiotic resistance. It's like if a high school worksheet got stopped at airport security.

Mythos Already Changed the Security Workload

Anthropic’s caution makes more sense when you look at what Mythos has already done.

In Project Glasswing, around 50 partners used Mythos Preview for defensive cybersecurity. Anthropic says they found more than 10,000 critical severity vulnerabilities.

Then Mythos scanned more than 1,000 open-source projects, initially estimated 23,019 vulnerabilities, and produced thousands of critical severity findings. Anthropic says independently reviewed findings had a 90.6% true-positive rate.

That’s the fear: the same model that can find vulnerabilities for defenders can also help attackers find them first. So Anthropic is keeping the public version on a much shorter leash.

That is the tradeoff with Fable 5. The model looks genuinely useful, especially for coding, but the launch is also a reminder that the best AI tools may increasingly come with routing, permissions, and a bouncer at the door.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico