What Claude Fable 5 actually does differently

Anthropic's line has three widely available tiers and one that most people will never touch. The one worth understanding properly is Claude Fable 5, because it is not simply "the big one" — it behaves differently enough that a system tuned for Opus or Sonnet will not automatically get the best out of it.

Where it sits

Per Anthropic's own model documentation, Fable 5 is described as their most capable widely released model, aimed at "next-generation intelligence for long-running agents". Opus 5 is the one they point you to for complex agentic coding and enterprise work. That distinction matters and is easy to misread: Opus 5 is not a cut-down Fable 5, it is the recommended default, and Fable 5 is the tier above it on raw capability.

Fable 5 costs twice what Opus 5 does on input and output alike. That is the number to keep in front of you for the rest of this piece.

Thinking is always on, and you cannot turn it off

This is the single most consequential difference. On Fable 5, adaptive thinking is always on. On Opus 5 and Sonnet 5 it is available; on Haiku 4.5 you get the older extended-thinking toggle instead.

The practical consequences:

• Latency has a floor. Anthropic classes Fable 5 as "slower" outright. If you are building anything interactive where a user waits on the response, this is a product decision, not just an infrastructure one.
• You cannot buy a cheap answer. On a trivial classification the model still reasons. Routing simple work to Fable 5 wastes money in a way that does not show up until the bill does — which is the argument for keeping Haiku 4.5 in the system for the small stuff.
• Prompts that manufacture reasoning are now counterproductive. "Think step by step before answering" is redundant at best. At worst you are spending output tokens duplicating a process the model is already running.

The tokeniser changed, and your budgets are wrong

Buried in the docs is a detail that catches people out. Fable 5 uses the tokeniser introduced with Claude Opus 4.7, and compared to models before that generation, the same text produces roughly 30% more tokens. The exact increase depends on content.

So the 1M context window is not directly comparable to the 1M window on Opus 4.6. Same nominal number, meaningfully less text. If you migrated a system that packed the window and you are now seeing truncation or unexpected cost, this is very likely why — and no amount of staring at your own code will reveal it.

Recheck any hard-coded token budget, any chunking threshold, and any cost model you built before the 4.7 generation.

Effort, and where its default changes

Anthropic exposes an effort parameter. The defaults are not uniform across surfaces: on Opus 5 and Sonnet 5 it defaults to high on the Claude API and Claude Code. If you want a different level you set it explicitly.

This is the most direct cost lever available and it is routinely left untouched. If your workload is high-volume and tolerant — summarising, extracting, triaging — a lower effort setting is the first thing to try, well before you consider changing model.

What the knowledge cutoffs imply

Fable 5 has a reliable knowledge cutoff of January 2026. Opus 5's is May 2026 — later than the more capable model. That inversion is worth internalising: the top tier is not automatically the best-informed one about recent events.

For anything where currency matters more than raw reasoning, that alone can decide the model. Better still, stop relying on parametric knowledge for current facts and give the agent a search tool, which is the correct answer regardless of tier.

Mythos 5, briefly

Claude Mythos 5 shares Fable 5's specs and pricing but is not generally available. It sits in Project Glasswing, invitation-only, for defensive cybersecurity work, with no self-serve sign-up. Fable is publicly described as the variant with the stricter safeguards. If you are reading a benchmark that includes Mythos, note that you probably cannot buy it.

A reasonable policy

• Haiku 4.5 for classification, routing and high-volume small calls.
• Sonnet 5 as the default agent loop. Most work belongs here.
• Opus 5 for agentic coding and long-horizon tasks where failure is expensive.
• Fable 5 when you have measured Opus 5 failing on the specific thing you care about, and latency is not part of the product promise.

Every tier here has a 1M context except Haiku, so "needs a big window" is no longer a reason to reach upward. Capability is the only reason left, and capability is testable. Test it.

If you want to try the Claude models yourself, you can sign up through this referral link.