# PointFive Labs: open AI efficiency research

Canonical: https://www.pointfive.co/labs

Meet Amir Hozez and Sarel Weinberger, the team leading PointFive Labs. Explore open research on AI infrastructure cost, with the papers, benchmarks, and data behind the findings.

Open, reproducible research on how AI infrastructure spend is actually incurred, with the benchmarks and data behind the findings.

[Meet the team](https://www.pointfive.co/labs#team) [Explore the research](https://www.pointfive.co/labs#open-research)

## The people doing the measuring

PointFive Labs is the research arm of PointFive, led by Amir Hozez and Sarel Weinberger.

Team announcement

### Introducing PointFive Labs

By Alon Arvatz, Co-founder and CEO

[Read the announcement](https://www.pointfive.co/blog/introducing-pointfive-labs) 

- Amir Hozez 

### Amir Hozez

Co-Head of PointFive Labs

Amir is co-founder and CTO of PointFive. Before starting the company he was VP of R&D at IntSights, which was acquired by Rapid7. He is a co-author of both Labs papers.

- Sarel Weinberger 

### Sarel Weinberger

Co-Head of PointFive Labs

Sarel has founded several AI companies and led AI at PwC. He lectures at Bar-Ilan University. He is lead author of both Labs papers and maintains the open benchmarks behind them.

## Token reduction is not cost reduction.

July 2026

[Read the research summary](https://www.pointfive.co/blog/token-reduction-is-not-cost-reduction)

Reducing what a coding agent sees does not necessarily reduce what it costs to complete the task.

The study compares Claude Code configurations using provider-billed cost and task outcomes. It separates delivered tool-output tokens from cache traffic and the additional work an agent may do after compression.

In the tested RTK-ML configuration, delivered tool-output tokens fell 38.4% while pooled billed cost rose 6.8%. These results describe the tested setup, not every compression tool or workload.

## Prompt-induced waste in coding agents

August 2026

[Read the research summary](https://www.pointfive.co/blog/prompt-induced-waste-in-coding-agents)

The same task can lead to different reasoning, verification, and costs depending on how the agent is asked to work.

The research examines prompt wording, inference effort, and the harness that manages tools and context. These factors interact: more effort can help difficult tasks, but can also add cost without improving outcomes.

Evaluation needs to track task success and end-to-end cost together, with the model, task, and harness kept explicit.

## Open benchmarks and data

Research tools

[Explore the benchmark and data](https://www.pointfive.co/blog/token-reduction-is-not-cost-reduction#benchmark-and-data)

Inspect the measurements behind the findings and reproduce the analysis.

The AI Efficiency Benchmark includes task manifests, retained campaign data, analysis scripts, and the methodology used to compare coding-agent configurations.

The benchmark records provider-reported cost and evaluates task success with deterministic checks. Its retained data can be re-analyzed without running a new paid campaign.

---
Source: the public page above. Product screenshots and illustrative interfaces are examples, not live customer data.