Breaking

Anthropic Redeploys Claude Fable 5, Scoring 80% on SWE-Bench Pro

July 3, 2026 at 19:32 EDT

Foundation Models
Software Dev & Coding
AI Agents

Anthropic released Claude Fable 5, its top-tier "Mythos-class" model, on June 9, 2026, and redeployed it on July 1 after a temporary suspension. The model targets software engineering and long-horizon agentic tasks, focusing on detecting technical debt and refactoring across large codebases.

June 9, 2026 · Anthropic

Claude Fable 5 leads the new "Mythos" class on autonomous coding

Anthropic's first generally available top-tier model foregrounds long-horizon software engineering — whole-repository reviews, technical-debt detection, and agentic tasks running for hours — at roughly double the price of Opus 4.8.

80.3%

SWE-Bench Pro — practical software engineering

token context, targeting multi-hour sessions

50M

lines of Ruby migrated in a single day (Stripe)

SWE-Bench Pro — score by model

Higher is better · Fable 5 leads the field

80.3%

Claude
Fable 5

69.2%

Claude
Opus 4.8

58.6%

GPT-5.5

The migration payoff — and its price

Stripe's full Ruby codebase migration, time to complete

2+ months

Traditional team

1 day

With Fable 5

Where it shines

Beats Opus 4.8 & GPT-5.5 on agentic coding
One PR review surfaced 16 findings — 3 real bugs
Months-long migrations compressed to days
Available via API, Claude Code, Copilot, Bedrock, Vertex

The catch

Review precision 32.8% — below Opus 4.8's 35.5%
High comment volume, noisy nitpicks
Guardrail routes to Opus 4.8 in <5% of sessions
Large reviews can burn tens–hundreds of dollars

API pricing — about 2× Opus 4.8

$10 / M

input tokens

$50 / M

output tokens

Jun 22

Pro/Max free period ends → metered billing

As the free window closes, cost versus benefit becomes the near-term question for users.

Continue reading

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