Anthropic just released Claude Fable 5, the most capable model they have ever made generally available, and it is now selectable across the CRHQ fleet. Here is what it is, what it is built for, how it performs, and when it is worth its price.
What Fable 5 is
Fable 5 is Anthropic's "Mythos-class" frontier model made safe for general use. Anthropic states its capabilities exceed any model they have made generally available before, with state-of-the-art results across software engineering, knowledge work, vision, and scientific research. It runs with a 1-million-token context window for big codebases and long sessions, and in CRHQ it appears as Fable 5 in the model picker.
What it is built for
Fable 5 is at its best on long-horizon, autonomous work: planning across stages, delegating to sub-agents, and verifying its own work. It can hold a long session without losing the thread, so it suits the jobs you would normally break into pieces.
- Root-cause investigations and outage debugging.
- Architecture decisions and large refactors or migrations.
- Deep, multi-stage research where extra reasoning pays off.
How it performs
On SWE-Bench Pro, the hardest agentic coding benchmark, Fable 5 leads the field by a wide margin.
How it behaves
- It thinks by default. Fable 5 decides how much to reason per step. In CRHQ it runs at xhigh effort, which favors thoroughness.
- It verifies its own work. You rarely need to remind it to test or double-check. Hand it the outcome you want, not step-by-step instructions.
- Built-in safety routing. A small share of requests, mainly in cybersecurity and biology, are automatically handled by Claude Opus 4.8 instead. This is expected and affects under about 5% of general sessions.
- Large context. A 1-million-token window holds big codebases and long-running sessions.
The cost, and when to use it
Fable 5 is the most powerful and the most expensive model we offer, roughly twice the cost of Opus 4.8.
| Model | Input / 1M | Output / 1M | Best for |
|---|---|---|---|
| Fable 5 | $10 | $50 | The hardest, longest, highest-value problems |
| Opus 4.8 | $5 | $25 | Complex reasoning |
| Sonnet 4.6 | $3 | $15 | Routine, everyday work |
| Haiku 4.5 | $1 | $5 | Light, high-volume tasks |
CRHQ tracks Fable usage and cost automatically, just like every other model. Prompt caching applies the usual discounts, with cache reads at 10% of input.
How to select it in CRHQ
- Pick Fable 5 in the model dropdown when starting a session, as an agent's default model, or in a background-job model override.
- Assign it to a specific agent, or switch to it for a single session.
- Everything else works the same as Opus and Sonnet: same interface, same tools.
- It is opt-in. No existing agent was switched to Fable automatically.
Tips to get the most from it
- Describe the outcome, not the steps. Give it the goal and let it plan the path.
- Hand it ambiguous, hard problems. Investigations, debugging, architecture, where extra reasoning pays off.
- Size up the task. Give it work you would normally split into pieces.
- Skip the "remember to test" reminders. It verifies more on its own.
Common questions
When should I use Fable 5 instead of Sonnet or Opus?
Use Sonnet for routine work, Opus for complex reasoning, and Fable 5 for the hardest, longest, highest-value problems where the extra capability is worth roughly double the cost of Opus.
How much does Fable 5 cost?
$10 per million input tokens and $50 per million output tokens, about twice Opus 4.8. CRHQ tracks the cost automatically.
Why was my Fable 5 request handled by Opus 4.8?
Fable routes a small share of requests, mainly cybersecurity and biology topics, to Opus 4.8 for safety. This is expected and affects under about 5% of general sessions.
Is Fable 5 the new default?
No. It is opt-in. Existing agents are unchanged until you select Fable for them.
Fable 5 raises the ceiling for what an agent in CRHQ can take on alone. Point it at your hardest problem, describe the outcome, and let it run.