The Joke That Stopped Being Funny to Me

If you depend on closed APIs like Claude (Anthropic) or ChatGPT (OpenAI) to run automations and critical processes, you’ve lived this nightmare: mid-shift, the service goes down, your team’s workflow freezes, and your customer is staring at a screen with no answer.

I’ve been watching Anthropic’s outages become a running joke on X and TikTok. At first, I laughed too. But at some point I stopped finding it funny, because I realized this instability isn’t just a temporary annoyance — it’s functioning as the best free commercial Open Source has ever had.

When the smartest proprietary model on the market goes down every other day, the question that keeps nagging at any software architect is simple: is it worth paying a premium for an intelligence that sometimes leaves you stranded?

The Boring Math of “Nines”

I’d never actually sat down to calculate what each “nine” of availability really costs, until I did the math.

  • 1 nine (98% uptime): the system is down about 2% of the time.
  • 2 nines (99% uptime): down 1% of the time.
  • 3 nines (99.9% uptime): the minimum standard any serious enterprise demands.

Today, while OpenAI tends to operate between 2 and 3 nines, Anthropic has racked up stretches that put it between 1 and 2 nines during parts of the month. From the outside, the difference between 98% and 99% looks like a rounding error. In engineering practice, it isn’t: a system with 1 nine is 5 to 10 times less stable than one with 2 nines. In an operation handling thousands of calls per hour, that turns into hours of downtime — and direct losses, no discount.

The “5 Nines” Open Source Hands You for Free

This is where open models — LLaMA, DeepSeek, Qwen — become almost unfairly compelling: continuous availability.

When you run an open model on your own infrastructure — a dedicated cloud server, or even a homemade cluster of Mac Minis — your reliability jumps to 5 or 6 nines (99.999%). As long as there’s electricity and your machine is on, the model is standing. No third-party server to crash, no API queue to freeze your operation.

And even if you’d rather not run your own servers and consume open models through providers like Perplexity, OpenRouter, or Together AI, redundancy is nearly instant: one provider goes down, you redirect traffic to another in seconds.

The Trade I Didn’t Expect to Do the Math On and Agree With

What caught me most off guard was the financial side of this migration. It’s not just about stability — it’s about money leaking out of the budget for a reason that stopped making sense.

Startups and tech companies are reporting 95% to 97% reductions in their monthly bill after switching from proprietary APIs to optimized open models. And the price of that trade, in quality, tends to be just 3 to 4 percentage points — because today’s leading open models already handle roughly 98% of real day-to-day tasks (engineering, support, analysis).

When I put that side by side, it’s hard to justify paying a fortune for a closed, unstable API just for 2% or 3% of extra intelligence.

CriteriaClosed proprietary model (SaaS Cloud)Open Source model (self-hosted / decentralized)
ReliabilityLow to medium — depends on a third-party serverExtremely high — up to 99.999% locally
Cost at scaleHigh and unpredictable — pay per tokenVery low — fixed infrastructure cost
Data privacyData travels through external serversAbsolute — data never leaves your environment
Intelligence level100% — top-tier reference95% to 97% of top-tier performance

What I Actually Think

I don’t think this means closed models are going away — Claude, when it’s up, still delivers a quality that many open models can’t match on harder reasoning tasks. But what this wave of outages is exposing is a truth we tend to forget when we’re dazzled by cutting-edge intelligence: whoever doesn’t control their own infrastructure doesn’t control their own business.

That’s not a new lesson. It’s the same one we already learned with cloud versus on-prem, with vendor lock-in, with any critical dependency handed off to someone else. AI is just re-teaching that lesson to a generation of companies that grew up assuming “reliable API” was a given by default.

What’s different now is that, for the first time, the cost of escaping that dependency has dropped low enough to be worth it — 3% less quality for 97% savings and real uptime is a math problem that pretty much solves itself.

I’m Left With This Question

I’m not saying anyone should ditch Claude or ChatGPT tomorrow. I’m saying that if your operation runs on a closed API with no backup plan, you’re already more exposed than you think — and the recent outages just made that visible to everyone at the same time.

Has your company ever gotten burned by an AI API going down in the middle of something important? Have you considered running an open model locally or on dedicated cloud to lock in real uptime?

97% savings for 3% less quality. When the math is this clean, the joke about Claude going down becomes the best sales pitch Open Source has ever had.

Read Also