Refresh
Refresh builds artificial worlds where AI agents are evaluated, trained, and improved through reinforcement learning. The company creates high-fidelity clones of software suites for computer-use testing and constructs 3D environments inside Blender for tasks like rigging, animation, and physics simulation. Refresh also publishes BlenderBench, a standardized benchmark of 20 Blender workflows with automated reward functions.
- Artificial Intelligence
- Developer Tools
- AI Agents
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
AI agents that operate software or work in 3D creative tools need realistic, controllable environments to be evaluated and trained effectively. Generic benchmarks and static datasets fail to capture the complexity of real workflows, making it difficult to measure agent capability, improve performance through reinforcement learning, or safely test models before deployment in production settings.
Solution
Refresh builds artificial worlds where AI agents are evaluated, trained, and improved through reinforcement learning. The company creates high-fidelity clones of full software suites—from electronic health records to enterprise applications—that wire together multiple connected applications for realistic computer-use testing. In the 3D space, Refresh constructs environments inside Blender where agents perform tasks such as building scenes, applying materials, rigging characters, keyframing animation, and running cloth and physics simulations. The company also publishes BlenderBench, a standardized benchmark of 20 Blender workflows with automated reward functions, providing measurable, repeatable evaluation of model performance.
Target Audience
Primary customers are AI labs and enterprises that need simulation environments to evaluate, train, and improve agent models for computer use and 3D creative workflows.
Features
- High-fidelity software world clones that replicate complete application suites and interconnect multiple apps for realistic computer-use agent training
- Blender-based 3D environments supporting scene construction, materials, character rigging, keyframed animation, and cloth/physics simulation tasks
- BlenderBench benchmark with 20 standardized workflows, each defining agent inputs, required changes, and elements to preserve, scored against a 0.90 reward threshold
- Automated reward functions and reviewed task results enabling quantitative comparison of agent performance across modeling, rigging, and animation task types
- Cost-performance tracking per task, with published inference cost data (e.g., $2.76 per task for GPT-6 Astra and $1.03 per task for Claude Opus 5.5)