Breaking

Anthropic's Claude Fable 5 Shows Sharp Benchmark Drop After Export Ban Lifted

July 2, 2026 at 14:54 EDT

Foundation Models
Policy & Regulation
Security

On July 1, 2026, Anthropic's flagship Claude Fable 5 returned from roughly three weeks of suspension tied to U.S. export controls, but the BridgeBench benchmark recorded a steep decline in performance. Operator BridgeMind suggests newly added safety mechanisms may be triggering excessively.

July 1, 2026 · Anthropic

Claude Fable 5 Returns — But Its Coding Scores Collapse

After an export-control suspension, Anthropic's redeployed flagship now falls back to a more conservative model even on ordinary coding tasks. Independent BridgeBench retests show debugging and refactoring performance falling off a cliff.

−60.3

Points lost in Debugging after redeploy

−35.2

Points lost in Refactoring

2×

Price of Opus 4.8: $10 in / $50 out per M tokens

BridgeBench: Before Suspension vs. After Redeploy

Score out of 100 — each column drawn to scale

86.2

25.9

Debugging

73.6

38.4

Refactoring

75.9

61.7

Halluc. resist.

Before suspension After redeploy

Why the Scores Dropped: The Timeline

JUN 9

Public launch as coding flagship

→

JUN 12

Export-control ban → suspended

→

JUL 1

Redeployed with new guardrails; falls back to Opus 4.8

Praise

Strong creativity for Remotion video generation and vibe coding
Still near the top for UI and Reasoning categories

Complaints

Guardrail fallback fires on routine debugging and refactoring
Tighter weekly limits exhaust allowances early
Large gap vs. the less-restricted Mythos 5

"The redeployed Fable 5 is not the same model as before" — benchmarkers argue new guardrails now trigger on ordinary tasks, quietly routing work to the more conservative Opus 4.8.

Safety classifier originally designed to fall back in <5% of sessions · Context up to 1M tokens · Max output 128k tokens.

Continue reading

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