Factory said on 29 October that Droid 1.0, the new engine behind its coding agent, is “more accurate, efficient, and faster than ever before”. On Terminal-Bench 4 with Claude Opus 5.5, Factory says Droid scored higher than Claude Code while taking 26% less time per passed task and using 17% fewer tokens.
The largest change, Factory says, is letting the model write a program that calls its tools. In one example from Factory’s own sessions, that cut the input the model read from an estimated ~3.1M tokens to ~61,000, about 51 times fewer by our arithmetic on Factory’s figures, with the larger number modelled.
Short explainer · Factory Droid 1.0
A vertical news Short narrated from the headlines and opening lines of this piece and two related pieces.
Business Pill · THE MODEL AND THE HARNESS
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and does not discuss Factory or Droid.
The key insight: Factory says its Droid 1.0 harness scored higher than Claude Code on Terminal-Bench 4 with Claude Opus 5.5 while using 17% fewer tokens, and that programmatic tool calling cut one task’s input to about 51 times fewer tokens, by our arithmetic on Factory’s figures with the larger one modelled.
Factory’s research post, “Droid 1.0: Building a Frontier Harness”, describes Droid as “our agent harness” and an agent harness as “everything around the model: the prompts, the tools, how tool results come back, and the loop that runs it all”.
Factory says it has worked “with hundreds of thousands of developers across some of the largest enterprises in the world” over the past year. It calls the update “the largest one yet” and says Droid 1.0 is available today as a desktop app for macOS and Windows, a terminal UI and an SDK.
Its post on X announcing the release had 292,268 views when we read it on 30 October.
Factory’s 29 October 2026 post gives before and after figures for Droid’s base prompt, launch and quit times, and one Linear cleanup task from its own sessions, where the before figure is Factory’s modelled estimate. After as a share of before is our arithmetic. These are Factory’s own figures, not independently rerun.
What Was Posted on X
Introducing Droid 1.0: the new engine behind Factoryโs agent.
Built for complex projects and everyday development: >more accurate >more efficient >faster than ever before
Factory’s own post announcing Droid 1.0 as the new engine behind its agent, built for complex projects and everyday development and described by Factory as more accurate, more efficient and faster than ever before.
The Head-to-Head Claims
Factory compared Droid with one rival harness at a time, on the same model, on Terminal-Bench 4, with three trials per task, run 23 to 28 October. Its methodology lists the rival versions it used: codex@0.163.0, claude-code@2.1.306, cursor-cli, devin and opencode.
With GPT-6.1 Sol, Factory says Droid “scores 3.5 points higher than Codex at 12% lower cost”, and 6.6 to 12.6 points higher than OpenCode and Devin. With Claude Opus 5.5, it says Droid scores 3.0 to 7.6 points higher than Cursor, OpenCode, Devin and Claude Code, “at up to 24% lower cost”.
On time, Factory says Droid on Opus 5.5 “takes 26% less time per passed task than Claude Code and uses 17% fewer tokens”. On GPT-6.1 Sol, it says Droid “uses 12% fewer tokens than Codex, while taking 10% longer”.
How the Engine Changed
Factory made programmatic tool calling the default: the model writes JavaScript that runs in a sandbox and can make hundreds of tool calls in one model request. Factory says this “cuts Droid’s cost by 15% on Claude Opus 5.5” compared with direct tool calling.
In a Linear cleanup from its own sessions, Factory says one Script moved 151 issues in 312 tool calls using 2 requests and ~61,000 tokens, against an estimated 36 requests and ~3.1M input tokens with direct calls. That is 18 times fewer requests and about 51 times fewer input tokens, by our arithmetic on Factory’s figures.
Factory also says its base prompt shrank 37%, from 20,887 to 13,055 input tokens, and that in production “the typical request reads 98% of its input from cache”.
Why It Matters for Coding Agents
Factory sells a harness that runs on other companies’ models, including those of Anthropic and OpenAI, which also ship their own coding agents. Factory says it is “model-independent by design” and that Droid “outperforms the native harnesses built for their models up to 20% more efficiently”.
As we read the post, Factory’s pitch is cost per passed task on the same model: it reports cost and time only for tasks both harnesses passed in all three trials, priced at each provider’s list price.
As we read it, the comparison puts Factory against the model makers’ own agents: Claude Code and Codex are among the five harnesses Factory says it tested Droid against.
As we read the methodology, the results carry Factory’s own adjustments, including Opus 5.5 trials rerun on Claude Opus 5 and some Devin tasks counted as passed where Codex passed.
Factory Research, Droid 1.0: Building a Frontier Harness, 29 Oct 2026
FOR TEAMS PAYING PER TOKEN Factory says Script mode cuts Droid’s cost by 15% on Claude Opus 5.5 compared with direct tool calling.
FOR THE MODEL MAKERS’ AGENTS Factory says Droid scores 3.5 points higher than Codex at 12% lower cost on GPT-6.1 Sol, and 3.0 to 7.6 points higher than Claude Code and three others on Claude Opus 5.5.
FOR BUYERS COMPARING AGENTS These are Factory’s own results on its own setup; the post gives point and percentage differences rather than absolute scores in the text we read.
The analysis puts it in one line: “The surface, not the model, governs autonomy.”
As we read it, Factory’s claim that the same model scores differently in different harnesses is one test of that idea. The framework was published in July 2026 and does not discuss Droid 1.0.
What Is Not Established
These are Factory’s own benchmark results on its own setup; we did not rerun them. The post gives point and percentage differences but, in the text we read, not each harness’s absolute scores.
Factory’s methodology says it counted only tasks both harnesses passed in all three trials, reran some Opus 5.5 trials on Claude Opus 5 after Anthropic’s content moderation stopped them, and counted some Devin tasks as passed where Codex passed. In the post we read, Factory does not give its revenue, a price for Droid or customer names.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Factory says Droid 1.0 scored higher than Claude Code on Terminal-Bench 4 with Claude Opus 5.5, taking 26% less time per passed task and using 17% fewer tokens, and scored 3.5 points higher than Codex at 12% lower cost on GPT-6.1 Sol. Its programmatic tool calling cut one task’s input from an estimated ~3.1M tokens to ~61,000, about 51 times fewer by our arithmetic.
A note on sourcing. All Droid 1.0 claims, benchmark results and methodology come from Factory’s research post of 29 October 2026, read in full on 30 October; they are Factory’s own results and were not rerun by FourWeekMBA. Figures labelled our arithmetic are our own.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
