Breaking

Anthropic's Claude Fable 5.1 Leads Opus 5 on Agentic Benchmarks, but Opus Stays the Default

September 3, 2026 at 00:46 EDT

Foundation Models
Software Dev & Coding
AI Agents

Anthropic's Claude Fable 5.1 now leads Claude Opus 5 on the company's published agentic benchmarks, but the firm still tells developers to start with the cheaper Opus 5 model and escalate only when that pass falls short.

Sep 1, 2026 · Anthropic

Fable 5.1 vs Opus 5

Anthropic's Mythos-class model leads Opus 5 on published agentic benchmarks at twice the list price. The company still tells teams to start with Opus 5.

Terminal-Bench-Science 0.1

Vendor-run agentic science scores. Column height is proportional to the published pass rate.

52.6%

Fable 5.1

29.0%

Opus 5

24.7%

Fable 5

22.4%

GPT-5.6 Sol

List price, input and output

Fable 5.1 still costs 2x Opus 5 on the sticker. Each block is $5 per million tokens.

$5 / $25

Opus 5 · 1x

$10 / $50

Fable 5.1 · 2x

Cache reads are the exception: Fable 5.1 is $0.25 vs Opus 5 at $0.50.

58%

fewer output tokens on Snorkel wins

36%

faster on median successful run

75%

cheaper cache reads vs Fable 5

66 vs 63

AA Intelligence Index at max effort

Where Fable 5.1 wins

Leads the vendor agentic table, including 52.6% vs 29.0% on Terminal-Bench-Science. Testers report fewer tokens when a run actually finishes, and some shops are moving hard jobs off Opus 5.

Where Opus 5 still holds

Half the list price, moderate latency, and more robust on Snorkel's matched set. Fable 5.1 can lose last-mile terminal work and can burn through usage limits on everyday loops.

Routing rule

Start with Opus 5. Escalate to Fable 5.1.

Use Opus 5 for most coding and agent loops
Raise effort before changing models
Move to Fable 5.1 when a high-effort Opus 5 pass still fails, or the job must run for hours

Continue reading

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