The day the agents all stopped

Issue #13 · September 5, 2026 · ~5 min read

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. On Tuesday, Wednesday and Thursday this week, three different labs each shipped a model built to work for hours without you. On Thursday morning, two of them stopped working — along with a third provider, inside the same few hours. Both halves of that are this issue.

Before we start, a reminder that I've started a Patreon. That's where the notes, the half-finished thoughts and the more personal videos that don't fit here or on the channel end up: patreon.com/ZerotoMVP

Now, let's go.

1. The signal of the week

Three labs shipped a "leave it running" model in three days. On the third day, two of them stopped working — and nobody will say why.

The launches came in a row, all selling the same change.

September 1 — Claude Fable 5.1. Base pricing unchanged at $10 / $50 per million tokens, but cache reads dropped 75%, to $0.25 per million — Anthropic's own estimate is around 25% off a typical workload. The customer quote it led with was MongoDB's: the model "ran for hours unattended… I would wake up in the morning to the next phase finished."

September 2 — Gemini 3.8 Flash, Google's third Flash model in six weeks, described in its own announcement as built for "long-horizon coding and autonomous agents" and as "our most intelligent workhorse model." A million tokens of input context at $0.75 / $3.75 per million — roughly a thirteenth of frontier price.

September 3 — GPT-6 Astra. OpenAI president Greg Brockman: "Welcome to the AGI era." On the launch tables it posts 57.7 on Terminal-Bench 4.0 against GPT-5.6 Sol's 37.3, and 72.6% on OSWorld 2.0 at roughly 40 minutes per task against Sol's 65.7% at about 75.

Read the three together and the pitch is identical: stop supervising, start specifying. The expensive part of the job moves to the spec, the decomposition and the done-criteria — which is fine, most working engineers already have that skill.

Then, at 8:07 a.m. Pacific on September 3, ChatGPT, Claude and Grok all reported outages on their own status pages inside the same window. Gemini spiked on Downdetector, but Google never confirmed anything. Full restoration came at 12:38 p.m. For scale: Claude's own 90-day uptime is 99.4%, ChatGPT's 99.63%.

Nobody has confirmed a common cause. SpaceXAI apologised publicly for "an outage at our Memphis compute center" and, notably, "to our impacted compute partners." Azure was the obvious guess, and Microsoft said that wasn't it. Days later there is still no explanation.

Here is what none of the launch posts mentioned. A 30-second call that fails is a retry. An eight-hour run that dies at hour six is six hours of tokens, an unknown amount of half-finished work, and no record of where it got to. Every one of these models is asking you to lengthen the interval between checks — and lengthening that interval is exactly what makes an outage expensive. Adding a second provider doesn't obviously fix it, either: you cannot see what your providers share until the day they fail together.

The conclusion isn't "don't run long jobs." It's that a long run needs what a short call never did: checkpointed state you can resume from, an artifact written at every phase boundary rather than at the end, and a cost ceiling that kills the run instead of a timeout that leaves it hanging.

Two caveats, both real. Every benchmark above is vendor-reported and run at maximum effort — and OpenAI quietly edited several of Astra's published figures hours after launch, competitors' numbers included. And the outage cause is genuinely unknown. I'm not telling you these providers share a datacenter. I'm telling you three of them failed together and none of them will say why, which for your purposes is worse.

2. My take

The competition between models has quietly changed subject: away from parameter counts and leaderboards, toward a better question — how much work can you hand over without checking?

That changes what "better" means. The classic agent failure isn't a wrong answer — it's that on step 20 the model slightly misreads the task, and by step 500 it's very efficiently doing the wrong thing. For anything running unattended, reliability at step 500 matters more than brilliance at step one, and that's what a benchmark table is worst at showing.

So I've stopped caring much about price per token. The number I want is the cost of finished work — and next to it, the one nobody publishes: how long it ran before a human had to step in.

GPT-4 answered questions. GPT-5 helped you work. This generation is the first that takes on work you can genuinely delegate — which makes "how many hours of human thinking can we hand over safely" the number to watch. I'd bet it moves faster over the next few years than anything on the launch tables.

3. What I'm testing this week

Fable 5.1's cache discount, treated as long-run economics.

Normally this block is something I've already used. This week it's something I'm about to, and I'd rather say that than dress an argument up as a finding.

Anthropic left base pricing alone and cut cache reads by 75%, to $0.25 per million. That reads like a footnote until you think about what a long agent run actually does: it re-reads the same repo map, the same spec, the same accumulated plan on every single step. A thirty-second call barely touches the cache. An eight-hour run touches almost nothing else.

So, the experiment this weekend: one real multi-hour job, run twice. Once with the prompt as I normally write it, once restructured so the stable part is cacheable — system prompt, repo map and spec pinned at the front, volatile state at the back. Same task, same model, same effort setting.

Two numbers I don't have yet: the delta on the bill, and whether the restructured prompt made the output worse. I'll publish both next week, including if the answer turns out to be "no meaningful difference." A promised measurement is only worth reading if the null result gets published too.

What I already expect to go wrong: the discount only exists if you actually hit the cache, and any edit near the front of the prompt invalidates everything after it — which is precisely what a naive agent loop does when it appends state in the wrong place. And this is a discount on one component, not a cheaper model. Base rates are unchanged, so a run that mostly generates rather than re-reads will see almost none of it.

4. MVP idea of the week

A provider-health prober that knows about your jobs.

Problem: on September 3 the status pages lagged the failure. If you had an agent mid-run, it kept spending tokens against a dead endpoint until something timed out. Status pages answer "is the provider up." Nobody answers the question you actually have: is the thing I have running right now still working, and should it stop?
Solution: a small always-on prober that sends one cheap call to each provider you actually use, on a fixed interval, and records latency and error class. The part no status page can do: it knows which of your own long-running jobs are in flight, and on a sustained failure it pauses them at the next checkpoint instead of letting them burn tokens into a wall.
For whom: anyone running unattended jobs longer than an hour — a group that got considerably bigger this week and has no operational habits yet.
Minimal v1: a cron entry, a SQLite file, one webhook. It checkpoints and pauses. It does not fail over — failover is a much harder product, and this week suggests it might not even help.
Demand check: run it across three providers for two weeks and publish the chart. On September 3 that chart would have been the story, and nobody had one. If people ask for the data rather than the tool, the data is the product.
Why now: the failure mode stopped being hypothetical on Thursday — and the lesson isn't "add a second provider," it's "know, in your own process, the minute it stops."

5. Useful links

collusion.wiki — 18,000 wiki edits no human wrote; the timeline, and the sandbox bypasses agents shared.
Anthropic — Formalizing Fermat's Last Theorem — eleven days, dozens of agents in parallel; read how the work was actually split.
ARC Prize — Astra — 62.7% or 99.9% on the same benchmark, depending only on the harness.
Fortune — Astra's evaluation metrics changed after publication — published numbers moved hours after launch, competitors' figures included.
Nvidia confirms it will buy Hugging Face — $12.93B, confirmed September 3; the open-model hub now has an owner.

— Nick

P.S. Five of my own things, on GitHub: