Factory Droid 1.0
Factory’s model-independent coding agent harness got its largest update on 31 October 2026: programmatic tool calling is now the default, so the model writes a sandboxed program that runs hundreds of tool calls while only printed output returns to the context.
Overview
One-line facts, source-labelled.
| Vendor | Factory |
|---|---|
| Form | Model-independent coding agent harness (desktop app, terminal UI, Node SDK) |
| Launched | October 31, 2026 (Droid 1.0; research post October 29) |
| ARscore | not scored |
| Pricing | bring any supported model or your own API keys |
| Core change | programmatic tool calling (Script mode) on by default |
| Open source | No |
Harness comparison (Terminal-Bench 4, Claude Opus 5.5)
Factory ran every harness against Droid on the same model.
Droid
at $4.73 per passed task
Cursor
same model, same runs
OpenCode
same model, same runs
Devin
same model, same runs
Claude Code
same model, same runs
Bar length equals Terminal-Bench 4 pass rate on Claude Opus 5.5. All figures self-reported by Factory (3 trials per task, runs 23–28 October 2026), not independently reproduced. Vendor self-claim
Second model check (GPT-6.1 Sol)
The same harness race replayed on an OpenAI flagship.
| Harness | Pass rate | Cost per passed task |
|---|---|---|
| Droid | 57.6 | $0.79 |
| Codex | 54.0 | — |
| OpenCode | 51.0 | — |
| Devin | 45.0 | — |
Same caveat: Factory’s own runs. Droid is 10% slower per passed task than Codex on this model, by Factory’s own numbers.
Features
What changed in the harness (presence, not benchmark performance).
| Capability | Status | Source |
|---|---|---|
| Script tool: model writes sandboxed JavaScript that runs the tools | ✓ | Official |
| Raw files, search results and logs stay inside the program | ✓ | Official |
| Independent tool calls run concurrently | ✓ | Official |
| Base prompt cut 37% to 13,055 input tokens | ✓ | Official |
| 2 tool definitions up front, down from 11 | ✓ | Official |
| Typical production request reads 98% of input from cache | ✓ | Official |
| Works with any supported model, including your own API keys | ✓ | Official |
| Factory Router picks a cost-efficient model mix | ✓ | Official |
Efficiency figures (vendor-reported)
Factory’s own sessions, shown as before/after.
| Task | Direct calls | Script mode |
|---|---|---|
| Unit-test fix (15 tool calls, GPT-6.1 Sol) | 10 model calls, 56.8 s | 5 model calls, 41.6 s |
| Linear cleanup (151 issues, 312 tool calls) | est. 36 requests, ~3.1M tokens | 2 requests, ~61,000 tokens |
| CI sweep (5,384 jobs across 547 runs) | — | 2 requests; under 0.1% of tool results reached the model |
Script mode cuts cost by 15% on Claude Opus 5.5 versus direct tool calling; on by default for 10 models including GPT-6.1 Sol and Kimi K3. Vendor self-claim
Setup and configuration requirements
What you need before first use.
| Runs on | macOS and Windows desktop app, a terminal UI, and the Node SDK @factory/droid-sdk |
|---|---|
| Model | any supported model or your own API keys; Factory Router can choose the mix |
| Script mode | on by default for 10 models; sandbox has no file, network or system access of its own |
| Deployment | runs locally and in the cloud; sessions share one daemon (~20 MB each) |
Use cases
Scenario coverage, not benchmark performance.
Bulk operations
Move and verify 151 tracker issues in two model requests.
CI sweeps
Thousands of jobs checked with almost nothing loaded into context.
Delegated development
Delegate work at scale; ship PRs, configured infrastructure, incident investigations.
Harness portability
Keep one workflow while swapping models — the model-independent pitch.
Download and official links
Outbound, for trial and verification.
| Release note | https://factory.com/news/frontier-harness |
|---|---|
| Benchmark page | https://factory.com/news/terminal-bench |
| Factory Router | https://factory.com/product/router |
| Factory | https://factory.com |
| News coverage | AgentRanks news feed |
Compare further
The sibling harnesses and the scored incumbents.