Claude Fable 5 just shipped. Here's what it actually means for your business.

Anthropic recently released Claude Fable 5 — the first of its Claude 5 family, and a new "frontier" tier that sits above Opus, which was previously the top of the range. (There's a sibling called Mythos 5, which is the same model offered to a small set of approved organizations under a separate programme — for everyone else, Fable is the top shelf.)

If you run a business rather than an AI lab, the question isn't "how good is it?" It's "does this change anything for me?" Here's the honest version.

The lineup now, in one table

As of this writing, the Claude range prices out like this (per million tokens, API pricing in USD — input / output):

• Haiku 4.5 — $1 / $5. Fast and cheap. Classification, routing, simple extraction, high-volume pipelines.
• Sonnet 5 — $3 / $15. The workhorse. Most production workloads — document understanding, assistants, drafting — live comfortably here, at near-Opus quality on a lot of tasks.
• Opus 4.8 — $5 / $25. The hard-problem tier. Complex reasoning, serious coding and agentic work.
• Fable 5 — $10 / $50. The new frontier tier. Twice Opus pricing.

That last line is the point most coverage buries: the frontier tier costs 10× the workhorse tier on input and output alike. That's not a criticism — it's a signal about what it's for.

What the frontier tier is actually for

Fable 5 is built for the work the previous generation couldn't reliably do: very long, autonomous, multi-step tasks — the kind where a single request can run for many minutes while the model plans, executes, and checks its own work. Its reasoning is always on; you can't turn it off. That's a feature for a complex overnight analysis job. It's a bug for a customer-facing chatbot that needs to answer in two seconds.

For most of the workloads we build for clients — extraction pipelines, internal document assistants, workflow automation, intake triage — the honest answer is that the workhorse tier was already good enough, and "good enough, verified against your real documents" beats "most powerful" every time you're the one paying the bill.

The fine print that matters if you handle sensitive data

Here's the part that belongs on this blog specifically. Fable 5 comes with conditions that are easy to miss and directly relevant to regulated and data-sensitive businesses:

It requires 30-day data retention. Fable 5 is not available under zero-data-retention arrangements — organizations whose API data-retention setting is stricter than 30 days simply can't call it; requests are rejected. If your compliance posture (or a client contract) depends on a zero-retention agreement with your AI provider, the frontier tier is off the table for that workload, full stop.

Read that again as a general lesson, not just a Fable fact: every cloud model comes with data terms, and the terms are part of the product. Capability headlines travel fast; retention clauses don't. Before any new model touches client data, someone in your organization should be able to answer "where does the prompt go, how long is it kept, and under what agreement?" — in writing. (If nobody can, that's exactly the gap our Private & Secure AI work exists to close — including the cases where the right answer is a model running inside your own infrastructure, where the retention policy is whatever you decide it is.)

It refuses more, by design. Fable 5 ships with stricter safety classifiers than earlier models, and a request can be declined mid-task. Anthropic provides fallback mechanisms so an application can hand the request to another model automatically — but that's engineering someone has to do. If a vendor is wiring the frontier tier into your workflow, ask them how refusals are handled. "It just works" is not an answer.

It's slower and costlier per interaction. Always-on reasoning means longer response times and more tokens per request. Latency-sensitive workflows — support desks, live intake, anything with a human waiting — are usually better served a tier or two down.

What to actually do this month

Nothing dramatic. A new frontier tier is not a rip-and-replace event; it's a repricing event. The sensible response:

1. Map your current AI workloads to tiers. For each one: what model runs it, what does it cost per month, and what's the cheapest tier that clears your quality bar? Most businesses have never done this exercise and are quietly overpaying — in the "wrong tier" sense more often than the "wrong vendor" sense.
2. Identify the one workload that was previously impossible. The frontier tier earns its 2× premium on tasks that failed on Opus — very long autonomous runs, genuinely hard multi-step analysis. If you have one of those, it's worth a controlled trial. If you don't, you've saved yourself the premium.
3. Ask the data question before the capability question. Retention, training use, region. In writing. Every model, every vendor, every time.

That's it. No panic, no FOMO purchase, no migration project.

If you'd like help with step 1 — an honest map of your workloads against what they actually need — that's exactly what the free AI Review is. Thirty minutes, no slides, and you'll leave knowing which tier each of your use cases belongs in, whether or not you ever work with us.