Keep pulling the thread on Matan Grinberg.
Due to past negative experiences with cloud vendor lock-in, enterprises are unwilling to commit to a single AI model provider like OpenAI or Anthropic.
Factory proactively returned just under $2 million in revenue to all its customers because the product was not good enough to create obsessed users.
The belief that co-designing an AI model and its harness makes them better together is false; a harness supporting multiple models will outperform a single-model harness.
When new Opus or GPT models were released, they initially performed better on the TerminalBench benchmark within Factory's Droid than in their native harnesses like Claude Code or Codex.
The open source model GLM 5.2 is performing at a very high level, to the point where half of Factory's internal token usage is now on open models.
The share of tokens going to open models on the Factory platform grew from less than 1% at the beginning of the year to a double-digit percentage currently.
When Factory was founded 3 to 3.5 years ago, the enterprise market was not ready for fully autonomous agents and was barely ready for GitHub Copilot.
Enterprises are wary of vendor lock-in from cloud providers like AWS and Azure, who have historically offered low-cost initial contracts and then increased prices by 10x at renewal.
Factory's value proposition is providing the performance of models like Codex or Claude Code without the vendor lock-in that comes from using the model labs directly.
Factory avoids vendor lock-in by ensuring that automations and artifacts created with its platform live within the customer's own codebase.
Factory's core operating principle is to "create obsessed customers," which is viewed as an output metric, in contrast to Amazon's principle of "customer obsession," which is an input metric.
Factory released its Droid CLI product on September 26th, 2025.