Breaking

Peter Gostev finds Claude Fable 5.1 stronger on dynamic 3D as single runs hit $65

September 2, 2026 at 22:27 EDT

Foundation Models
Software Dev & Coding

Peter Gostev of Arena.ai has put Anthropic's newly released Claude Fable 5.1 Max through a hard-prompt bake-off against Fable 5, Claude Opus 5, GPT-5.6, Kimi K3, GLM 5.3, Qwen 3.8, Grok 4.6, and DeepSeek, finding a real quality step-up on dynamic 3D even as a single generation swung from a few cents to $65.

Sep 2, 2026 · Arena.ai

Fable 5.1 is richer. The bill still stings.

Peter Gostev ran Claude Fable 5.1 Max against eight frontier models on 3D, SVG, games, and research prompts. Quality rose. One Westminster scene still cost $65.

$65

Westminster 3D on Fable 5.1 Max

$0.61

Kimi K3 on a comparable 3D prompt

75%

Cache-read price cut vs Fable 5

AA Intelligence Index at max effort

Same prompt, two-order cost swing

Column height is proportional to dollar cost. The cents-tier bars are a sliver next to Fable 5.1 Max.

$0.21

GPT-5.6 SVG

$0.61

Kimi 3D

$22

Fable SVG

$35

Venice game

$65

Westminster

Takeaway: The same SVG prompt ran $22 on Fable 5.1 and 21 cents on GPT-5.6 — about 100 times the price for extra detail.

Terminal-Bench-Science 0.1

Each block is about 5 percentage points. Fable 5.1 more than doubles Fable 5.

52.6%

Fable 5.1

29.0%

Opus 5

24.7%

Fable 5

22.4%

GPT-5.6 Sol

Where it gains

Richer, more dynamic 3D and a more realistic visual style. Science-agent score more than doubles Fable 5. Cache reads are 75% cheaper.

Where it does not

Not a clean sweep: Fable 5 still won a game comparison. Max runs stay expensive, and cheaper models can be good enough.

The upgrade is real. The default choice is still a tradeoff.

Max vs High is not a perfectly fair matchAA: 20% more per task at maxOpen-weight is not automatically cheap

Continue reading

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