Fable 5 Just Shipped: What Anthropic's Newest Model Means for Developers
Anthropic released Claude Fable 5 on June 9, 2026. The model sits above Opus and belongs to the Mythos family.
I have used the Claude API for over a year. This release forced me to re-read my entire prompt stack.
Technical details
- 1 M token context window
- 128 K max output
- Knowledge cutoff: January 2026
- Cost: $10 per million input tokens / $50 per million output tokens
Fable 5 costs double what Opus 4.8 costs, so treat it as a specialty model—use it only when errors are costly.
Three tiers
- Haiku 4.5: routing, classification, cheap extraction
- Opus 4.8: default for app traffic and coding
- Fable 5: complex, long-horizon tasks where accuracy matters most
Key changes
- Adaptive thinking is always on; you cannot turn it off.
- Trying to disable thinking or setting temperature triggers a 400 error.
- You must stream responses; large outputs will time out otherwise.
Security professionals should note a hidden risk: Fable 5 enforces hard safeguards for biology, chemistry, and cybersecurity. When a prompt hits a safeguard, the model falls back to Opus 4.8 instead of refusing outright. This can silently downgrade reasoning, which matters for smart-contract audits.
How to handle it
- Log the
response.modelfield on every call. - Verify the returned model matches your request.
- Don’t assume you received Fable 5 reasoning for every prompt.
I keep Opus 4.8 as my default and only opt into Fable 5 for the hardest tasks.
Watch the fallback behavior. A security model that changes its output on sensitive prompts is a risk. Read the model field. Log it. Never assume.