Factory Droid 1.0 makes AI coding agents faster and cheaper with new 'programmatic' approach

What you could do with it

• A developer could use Factory Droid 1.0 to fix unit tests or clean up code across hundreds of issues by writing a single JavaScript program instead of sending each tool call separately, cutting costs and time ¹.
• A development team could use Factory's built-in router to automatically pick the most cost-efficient AI model from a list that includes GPT-6.1 Sol and Kimi K3 for their coding agent tasks ¹.
• A researcher could run Factory Droid 1.0 on benchmark tests like Terminal-Bench 4 to compare how well its programmatic tool-calling approach measures against other coding agent harnesses ².

What changed in Droid 1.0

Droid 1.0 is the biggest update yet to Factory's coding agent system. The main change is that programmatic tool calling is now turned on by default for 10 supported models, including GPT-6.1 Sol and Kimi K3.

In a normal setup, every time the AI needs to run a tool — like searching for code or running a test — the result goes back into the model's memory. With programmatic tool calling, the model writes a JavaScript program instead. That program runs all the tool calls at once inside a protected sandbox. Only the final printed output returns to the AI. A single request can now drive hundreds of tool calls.

Factory says the idea builds on earlier research called CodeAct ¹². The company shared results from its own internal tests showing big savings. In one task fixing unit tests, 15 tool calls that previously took 10 model requests were cut down to 5 model calls, saving about 15 seconds. In another task cleaning up code across 151 issues, the old way used 36 requests and 3.1 million tokens; the new way used just 2 requests and about 61,000 tokens .

How Droid scores against competitors

Factory tested Droid 1.0 on Terminal-Bench 4, a benchmark that measures how well coding agents solve real terminal tasks. Each harness was run three times per task from October 23 to 28, 2026, using the same model ¹².

With Claude Opus 5.5, Droid scored 65.2%, beating the native harnesses from Cursor (62.1%), OpenCode (62.1%), Devin (59.6%) and Claude Code (57.6%). Droid also cost up to 24% less and ran 16% to 26% faster per passed task ¹².

With GPT-6.1 Sol, Droid scored 57.6% at $0.79 per passed task, edging out Codex (54.0%) and OpenCode (51.0%). However, on this model, Droid took 10% longer per passed task than Codex ¹².

Other improvements

Factory also made several behind-the-scenes changes. The base prompt shrank by 37%, from 20,887 to 13,055 input tokens. Fewer tool definitions are loaded at the start — down from 11 to 2 — with rare tools loading only when needed ¹².

The app loads faster, with first paint dropping from 1,839 milliseconds to 488 milliseconds. Each new session now uses about 20 MB of memory instead of 350 MB ¹².

Where to get it

Droid 1.0 is available as a desktop app for macOS and Windows, a terminal interface, and a Node SDK called @factory/droid-sdk. Paid plans start at $20 per month. The system works with any supported model, including the user's own API keys, or Factory's auto model router ¹².

Corrections and updates

• Corrected on 1 November 2026. An earlier version of this story included details its sources do not support. They have been corrected or removed.

Key facts

• Droid 1.0 scores 65.2% on Terminal-Bench 4 with Claude Opus 5.5, the highest among tested harnesses
• Programmatic tool calling is now default, letting one request run hundreds of tool calls
• Base prompt shrunk 37% — from 20,887 to 13,055 input tokens
• Only 2 tool definitions ship upfront, down from 11
• Pricing starts at $20/month for desktop, terminal, and SDK access

Sources

1. Meet Factory Droid 1.0: A Model-Independent Coding Agent Harness Built on Script Mode — marktechpost.com
2. Droid 1.0: Building a Frontier Harness | Factory