Breaking

Anthropic Restores Claude Fable 5 Worldwide, Routing Some Coding to Opus 4.8

July 1, 2026 at 02:05 EDT

Foundation Models
Software Dev & Coding
Policy & Regulation

Anthropic said it will again offer its top-tier model "Claude Fable 5" worldwide after an export-control directive forced a temporary suspension. As part of the restoration, the company added new classifiers that automatically fall back certain coding and debugging tasks to the previous-generation Opus 4.8.

July 2 · Anthropic

Claude Fable 5 Returns Worldwide — With Tighter Safety Guardrails

After a brief pause over jailbreak and misuse concerns, the flagship "Mythos-class" model is back online. Hardened classifiers now more strictly block cybersecurity-related queries, while only a small fraction of routine coding tasks fall back automatically to Opus 4.8.

95%+

of sessions run on full Fable 5 performance

<5%

average sessions routed to Opus 4.8 via classifiers

token context window, text · image · file input

Benchmarks: Fable 5 vs Opus 4.8

Higher score = better. Fable 5 leads on agentic & hard coding tasks.

SWE-Bench Pro

80.3%

Fable 5

69.2%

Opus 4.8

FrontierCode Diamond

29.3%

Fable 5

13.4%

Opus 4.8

How the fallback works

Sensitive queries are auto-detected and routed away from Fable 5.

User query

→

Safety classifier
cyber · bio · chem · distillation

→

Fallback to Opus 4.8
user notified

Strengths

Beats Opus 4.8 on multi-file migrations & autonomous coding
Stripe migrated a 50M-line Ruby codebase in a single day — work said to take a team 2+ months
Praised for planning, architecture & code review

Limits & Complaints

Classifiers can overreact — even to defensive security audits
Roughly 2× the price of Opus 4.8; costly for all everyday work
Gap narrows on pure coding; can "fail confidently"

Input price

$10 / M tokens

Output price

$50 / M tokens

vs Opus 4.8

~2× the cost

Continue reading

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