What Liberty Mutual knew that most companies didn't

THE AI BRIEF

Today's signal: The Fable 5 outage had a winner. Liberty Mutual kept running because they'd built for exactly this scenario. Here's the minimum viable version of what they did.

In today’s issue:

• Main story: The Fable 5 outage had a winner. It wasn't a tech company.

• Also worth knowing: Sysdig documented the first fully autonomous agentic ransomware attack, Grok 4.5 costs one-eleventh of Fable 5 Max per coding task, Cognition's SWE-1.7 runs at 1,000 tokens per second on Cerebras, and more

THE READ

When Anthropic pulled Fable 5 from international use for nearly three weeks, Liberty Mutual pivoted without pausing because they had built for exactly this kind of scenario.

When Anthropic pulled Fable 5 from international use last month, most of the coverage focused on the capability loss: developers losing coding tasks, enterprises scrambling for fallbacks, the gap between what Fable promised and what Fable's government-mandated guardrails delivered. That was the real story for builders. The enterprise story was quieter.

VentureBeat reported that Liberty Mutual's AI team essentially shrugged. When Fable 5 went down, they rerouted to other providers in the normal course of operations. What they'd built was a model-agnostic AI backbone, an infrastructure layer that treats individual model providers as interchangeable components rather than load-bearing dependencies, so when one component became unavailable, the system kept running.

What I keep hearing from operators who went through the Fable 5 disruption is a version of the same question: how do you build AI that doesn't depend on any one vendor staying available? My honest answer is that the full solution is more work than most teams are ready for, but the minimum version is simpler than they think. The difference that matters is between building on a model and building on a platform you control. A team that builds directly on GPT-5.6 or Fable or Claude has done the easy version of the work. The integrations are faster, the prompts are simpler, the initial results are often better. But they've also handed a single vendor a veto on their operations, and when the model gets pulled, the pricing changes, or the terms shift, you're negotiating from a position of dependency.

Get your team’s AI readiness score. Leave with an AI readiness brief.

A 15-minute assessment that turns your answers into a real decision artifact: your readiness level, your six-axis shape, where you're strongest, where you're constrained, and what not to build yet.

Liberty Mutual's approach is more expensive to build initially. You have to abstract the model layer, which means more engineering upfront, more testing, and more interface standardization. What you get back is an AI program that continues to function regardless of what any single vendor does, which is increasingly worth the cost.

The honest counterargument is that model abstraction introduces its own complexity, and for most mid-market companies, it is probably overkill in 2026. You're not Liberty Mutual. You have fewer systems, fewer integrations, and less exposure if a single model goes down for a few weeks. Most disruptions are temporary. The cost of building a fully model-agnostic infrastructure may not be justified by the risk.

That's true. But there's a version of this that doesn't require building Liberty Mutual's entire stack. The minimum viable version is knowing, right now, which of your AI workflows would pause if your primary model became unavailable, and having a tested fallback for the ones that would hurt. Not a full abstraction layer. A list and a backup. Most companies I talk to don't have either.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

ALSO WORTH KNOWING

Sysdig documented JADEPUFFER, the first fully autonomous agentic ransomware attack. An LLM agent exploited a Langflow vulnerability, pivoted to a production MySQL server, adapted when steps failed, and conducted credential sweeps across 600 self-narrating steps with no human operator. Any team running AI agents with database or file system access needs a security review on what those agents can reach.

The 2026 AI Engineering Survey (1,000+ respondents) from AI Engineer, Notion, and Vercel is out, covering model selection, build vs. buy decisions, and who is actually shipping AI in production versus still piloting. If you're making tooling decisions in the next quarter, this is the most comprehensive practitioner-reported dataset available right now.

Grok 4.5 costs $1.51 per coding task on CursorBench vs. $17.32 for Fable 5 Max, roughly one-eleventh the cost for 3.8 percentage points less accuracy at number three overall. For teams running high-volume coding workflows that don't need top-of-benchmark performance, Grok 4.5 is the first credible cost argument against the premium tier.

Cognition released SWE-1.7, their most capable coding agent yet, scoring within a few points of frontier models while running at 1,000 tokens per second on Cerebras. Free for all paid Devin users for the next month. The speed improvement matters more than the benchmark: agents that return results in seconds fit into developer workflows; agents that take minutes create friction that limits adoption.

Senator Bernie Sanders attacked Microsoft publicly for cutting 3,200 Xbox jobs and raising console prices by $150 while reporting $101 billion in profits and receiving a $12.5 billion tax break. The political pressure on companies cutting headcount while citing AI investment is building, and this is the first major Congressional-level public attack on the AI/employment story.

WATCHING TOMORROW

The AI Engineering Survey data drops Friday morning fully; the model selection numbers in particular will be worth a close read ahead of the Week in AI issue. Also watching: Anthropic's Fable 5 extended access window runs through July 12, and any change before then would reopen the vendor dependency conversation.

REPLY

Do you know right now which of your AI workflows would pause if your primary model went dark for two weeks? I read every reply.

If you know an operator running AI in production who hasn't thought through model dependency, forward this issue to them.

Back tomorrow,
Haroon