Breaking

Anthropic Publicly Releases Safety-Gated Top Model Claude Fable 5

July 4, 2026 at 18:05 EDT

Foundation Models
Software Dev & Coding
AI Agents

On June 9, 2026, Anthropic released Claude Fable 5 (API model ID: claude-fable-5), a general-availability model that adds safety guardrails to its most advanced Mythos-class. It touts greater autonomy on long-horizon, complex tasks than the prior flagship Opus 4.8, while automatically switching to Opus 4.8 for sensitive topics.

June 9, 2026 · Anthropic

Claude Fable 5 Opens Long-Horizon Autonomy to a General Audience

A new top "Mythos-class" model built for autonomous work spanning days to weeks — not single prompts — surpassing the former flagship Opus 4.8, with a safety classifier that falls back on sensitive topics.

80.3%

SWE-Bench Pro (vs Opus 4.8: 69.2%)

<5%

Sessions triggering safety fallback

50M

Line Ruby codebase migrated by Stripe in one day

SWE-Bench Pro: real-world software engineering

Same unit (% solved) — column height ∝ score

69.2%

Opus 4.8

80.3%

Fable 5

Pricing (per million tokens)

Input$10

Output$50

Prompt cache discount−90%

Roughly double Opus 4.8; caching cuts input cost sharply.

Terminal-Bench 2.1 (agent eval)

88.0% Mythos 5 equivalent (guardrails removed)

~84.3% Fable 5, lowered by safety refusals

Also SOTA-class on FrontierCode, GPQA Diamond, OSWorld, vision & financial-analysis benchmarks.

What developers praise

Deep reasoning & planning like Opus, execution efficiency like Codex
Beats Opus 4.8 on bug-finding, optimization, math rigor, multi-step planning
"Fable solved a task Opus gave up on"

Where it falls short

High price & heavy token use, especially under sub-agents
Slow, verbose comments; refusal rates can exceed 20% by setup
"Superior to Opus but sometimes inferior to Codex"

Recommended workflow

Fable 5

Design decisions

→

Opus 4.8

Execution

→

Sonnet / Haiku

Light work

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.