Breaking

Claude Fable 5 Hits 90.4% on Vibe Code Bench

June 10, 2026 at 07:08 EDT

Foundation Models
Software Dev & Coding
AI Agents

Claude Fable 5, the top-tier model Anthropic released publicly on June 9, 2026, scored 90.4% on Vals AI's full-stack app-building benchmark, Vibe Code Bench. Just six months ago, no model could break 20% on the same measure, marking a steep jump in capability within a matter of months.

June 9, 2026 · Anthropic

Claude Fable 5 Hits 90% Where No Model Could Crack 20% Six Months Ago

On Vibe Code Bench — which tests whether an AI can build a fully working web app end-to-end — Fable 5 scored 90.35%, the first model to reach the 90s and the new leaderboard leader.

90.35%

Vibe Code Bench score (±2.10) — top of the leaderboard

<20% → 90%

Leap on this benchmark in roughly half a year

Token context window · up to 128k output

Vibe Code Bench — building full web apps from a spec

Same scale (0–100%). Column height is proportional to the score.

90.35%

Claude Fable 5

82.72%

Claude Opus 4.8

69.85%

GPT-5.5

Verified across 964 workflows and 10,131 substeps in a browser — apps must actually function, not just compile.

80.3%

SWE-Bench Pro

88.0%

Terminal-Bench v1.1

What developers praise

Ahead of GPT-5.5 on long-horizon tasks and complex planning
Strong code quality, creative ideas — and a leap as a coding agent
"Best yet for creative writing and rap generation"

What raises concern

Slow and expensive; roughly double GPT-5.5's price
Occasional wasteful code, e.g. omitted return statements
Strict guardrails lead to more refusals

Pricing · per million tokens

$10

input

$50

output

No surcharge for long context · January 2026 knowledge cutoff · the GA, safety-tuned version of the Mythos class.

Continue reading

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