Breaking

Anthropic Reinstates Fable 5 With Updated Safeguards, Retaking Agent Arena Lead

July 1, 2026 at 16:56 EDT

Foundation Models
AI Agents
Software Dev & Coding

Anthropic said on July 1, 2026 that it had restored access to its large language model "Claude Fable 5". The model, which had been suspended on June 12, returned under updated safety guardrails and reappeared on the leaderboard of Agent Arena, a benchmark that evaluates real-world tasks.

July 1, 2026 · Anthropic

Claude Fable 5 Returns — and Reclaims the Top of Agent Arena

After a three-week suspension over export controls and security concerns, Anthropic restored access to Fable 5 with updated safeguards. It immediately retook first place in the benchmark for real-world, long-horizon agentic tasks.

13.34%

Net Improvement — #1 across 16,082 sessions

88%

FrontierMath Tier 4 benchmark score

1.31%

Tool Hallucination — lowest of the field

Net Improvement — Agent Arena leaderboard

Relative advantage in real-task outcomes. Higher is better.

13.34%

Claude Fable 5
(High)

9.37%

Claude Opus 4.8
(Thinking)

8.21%

GPT 5.5
(xHigh)

The suspension & return timeline

Three weeks from launch to restored access.

Jun 9

Made publicly available

→

Jun 12

Access suspended; removed from leaderboard

→

By Jun 30

Controls lifted; access restored with new safeguards

→

Jul 1

Back at #1 in Agent Arena

Strengths in real use

Confirmed Success rate: 16.12%
Praise vs Complaint: 30.63%
Wide lead in Code Arena: Frontend
Excels at long, multi-step workflows

Open concerns

Steerability weak at 9.21% — hard to redirect
Possible capability degradation from new safeguards
Usage capped at 50% of weekly limit through Jul 7

Continue reading

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