Measured by us

Agents and tool use — Claude Opus 5.5 on AWS Bedrock (us-east-1)

Results

Task success rate	84.5% higher is better
Median steps	7 steps lower is better
Error recovery	78% higher is better

Run details

Model version: claude-opus-5-5
Endpoint: Bedrock Mantle Messages API, effort=high
Platform / region: AWS Bedrock / us-east-1
Dataset: tool-tasks-200 (200 multi-step tasks, 12 tools)
Harness: GitHub @ a1b2c3d
Run cost: $41.70
Date: 23 Sept 2026

What was NOT tested

Computer use
Tasks longer than 50 steps
Parallel sub-agents

Re-run it on your data

git clone https://github.com/mallimatla/uptodate.git && cd uptodate && git checkout a1b2c3d
npm install && npm run proof -- --suite agents --model claude-opus-5-5 --platform bedrock --region us-east-1 --dataset ./your-data