Claude Fable 5.5 (Leaked Checkpoint) vs ConvAI Laya
A direct, empirical evaluation of Claude Fable 5.5 (Leaked Checkpoint) (Anthropic) against ConvAI Laya (ConvAI Innovations (Vishal Mysore / visrow)).
Claude Fable 5.5 (Leaked Checkpoint)
Developed by Anthropic · Proprietary Frontier Dynamic Reasoner parameters
Terminal-Bench 4.0 autonomous shell execution (78.5% target), recursive self-critique, and 1.5M-token repository refactoring.
ConvAI Laya
Developed by ConvAI Innovations (Vishal Mysore / visrow) · 421 Million (ModernBERT-large Backbone) parameters
Client-side in-browser decision execution via WebGPU/WASM, zero-cost operational routing, private on-device medical/financial triage, and domain-specific fine-tuning.
| Metric | Claude Fable 5.5 (Leaked Checkpoint) | ConvAI Laya |
|---|---|---|
| SWE-bench Verified | 84.2% | 64.2% |
| MMLU-Pro | 94.6% | 82.4% |
| MATH-500 | 98.4% | 88.5% |
| Context Window | 1,500,000 tokens | 8,192 tokens |
| Input Pricing (per 1M) | $3.50 | $0.00 |
| Output Pricing (per 1M) | $17.50 | $0.00 |