Factory, the company behind the Droid coding agent, released Droid 1.0 on 29 October and claims it now outperforms the tools that OpenAI and Anthropic ship for their own models, at lower cost. The claim rests on a benchmark Factory ran itself, but the mechanism behind it is worth understanding, because it changes how much a long coding job costs.

Start with the word “harness.” A coding AI model on its own only predicts text. The harness is the software wrapped around it that makes it useful: the instructions it receives, the tools it can use to read files and run commands, how tool answers are fed back, and the cycle that repeats until the job is finished. Codex and Claude Code are harnesses. So is Droid, which Factory built to work with many models instead of one.

The biggest change in version 1.0 is how the model uses those tools. Normally the AI asks for one action, reads the answer, then asks for the next, and every round trip resends the full instructions and the growing history. Factory now has the model write a short JavaScript program that makes all the calls itself and reports back a summary. The program runs in an isolated sandbox that cannot touch files or the network directly. Factory calls this programmatic tool calling, and says the idea comes from a 2024 research paper called CodeAct.

Factory’s own examples show where the savings come from. A cleanup in its Linear project tracker shifted 151 issues from one queue to another and checked each one, 312 tool calls in total. By Factory’s estimate, ordinary calls would have needed 36 model requests and about 3.1 million input tokens. The program needed 2 requests and about 61,000 tokens. A second job checked 5,384 test jobs across 547 runs, and the model only ever saw a short summary that flagged 41 failures.

Across a full benchmark, the gains are smaller but still consistent. On Terminal-Bench 4, Factory reports that Droid with GPT-6.1 Sol scored 3.5 points above Codex at 12 percent lower cost. With Claude Opus 5.5, the lead over Cursor, OpenCode, Devin, and Anthropic’s Claude Code ran from 3.0 up to 7.6 points, with costs up to 24 percent lower. The script approach alone cut Droid’s cost by 15 percent on Opus 5.5.

Speed is mixed. With Opus 5.5, Factory says Droid finishes each solved task 26 percent faster than Claude Code while using 17 percent fewer tokens. On OpenAI’s GPT-6.1 Sol, Droid spends 12 percent fewer tokens than Codex yet runs 10 percent slower.

The benchmark was run by the vendor, and its methodology notes show judgment calls. Devin cannot disable its habit of pausing to put a question to the user, so on GPT-6.1 Sol tasks where it paused, Factory gave Devin credit whenever Codex had solved the task, which happened on 4.7 tasks per trial on average. Factory says the adjustment touches accuracy scores only, leaving cost and time alone. No independent group has reproduced the numbers.

Smaller changes round out the release. Factory cut the starting prompt by 37 percent, from 20,887 to 13,055 tokens, mostly by sending 2 tool definitions up front instead of 11 and loading the rest on demand. It also says the typical production request now reads 98 percent of its input from the provider’s cache, which matters because cached text costs less to process.

The commercial point is model independence. Developers who want to switch between Anthropic, OpenAI, and open-weight models like Kimi K3 usually lose the polish each lab puts into its own harness. Factory is betting that a neutral harness tuned for every model can close that gap. A 12 percent price difference against Codex is small enough that your own codebase could erase it, so it takes about an afternoon of the same tasks on both to find out.

Reported by Factory on 29 October 2026, in its post “Droid 1.0: Building a Frontier Harness.”