Droid 1.0 squeezes 312 tool calls into 2 requests, but commenters zero in on the desktop app

Factory released Droid 1.0 on October 29. Matan Grinberg's first sentence on X was blunt: on Terminal-Bench, FactoryAI beats the rest. Droid is Factory's agent harness. A harness is everything outside the model: prompts, tools, how tool results are returned, and how the loop runs. The model handles reasoning; the harness is responsible for keeping the reasoning process from burning through tokens and context.

The Terminal-Bench 4 comparison used the same model and the same tasks. With GPT-6.1 Sol, Droid costs $0.79 per passed task, with a 57.6% pass rate. Codex is 54.0%, OpenCode 51.0%, and Devin 45.0%. Droid is 3.5 percentage points higher than Codex and 12% cheaper. With Claude Opus 5.5, Droid is $4.73 and 65.2%. Cursor and OpenCode are 62.1%, Devin 59.6%, and Claude Code 57.6%. Droid is 3.0 to 7.6 percentage points higher and up to 24% cheaper.

Note that these numbers are cost per passed task and pass rate, not per-request price. The longer an agent runs, the more errors compound into later steps, and high cost also limits parallel trial and error. The benchmark ran from October 23 to 28, with three trials per task; failures were rerun outside the harness.

The biggest change in Droid 1.0 is that programmatic tool calling is now the default. In an ordinary agent loop, each time the model calls a tool, the result is appended to context. PTC lets the model write a program; the program calls tools in a sandbox, so one model request can run dozens to hundreds of tool calls.

Factory gave a Linear example: moving 151 issues and verifying each one, for 312 tool calls total. The old approach took about 36 requests and roughly 3.1M input tokens. After switching to a Script, it took 2 requests and about 61K tokens. Another CI scan example checked 5,384 desktop jobs across 547 runs, with 550 tool calls, completed in 2 requests. The model only saw a five-line summary, which included 41 failed jobs; the raw results stayed in the program.

The reason PTC works is not complicated: raw files, search results, and logs are no longer all stuffed into the model context; loops and validation run deterministically in code; planning is moved forward to the moment the program is written. The gains mainly come from cutting context and request count: fewer repeated requests, and raw results no longer pile up in the model context.

There are plenty of engineering details too. Tool discovery used to be inserted at the top of the request, invalidating the prompt cache prefix. Droid now appends new tools to the end of the request. In production data, 98% of input for a typical request is read from cache, and 98.6% of requests are strictly append-only. The base prompt dropped from 20,887 tokens to 13,055, a 37% reduction. Tool definitions dropped from 11 to 2, cutting tokens by 45%. Startup reminders dropped by one third, while the system prompt actually grew by 31%, used for communication and environment details.

Startup and exit were also changed. First paint dropped from 1,839 ms to 488 ms, and time to accept the first prompt dropped from 1,876 ms to 618 ms. Median exit time dropped from 648 ms to 30 ms. Multiple sessions share a daemon, adding about 20 MB per session, instead of about 350 MB per session.

These changes come from Factory's internal Agent Factory. New models and snapshots arrive every few weeks; internal agents read failure traces, find recurring causes, propose changes, and run controlled experiments on every affected model family. Engineers look at the evidence and decide whether to merge. One example is adding line numbers when reading files: it stopped Sol from rereading files through the shell just to get line numbers; on one task, Opus cost dropped 22%, but GLM cost rose 8% and Kimi rose 32%. In the end, it was made an option that the model requests only when needed.

The X comment section tells another story. Avery Jones said the efficiency jump from 3M tokens to 60k is absurd. Others joked about the 'second tier'; Raja Rao DV asked whether Factory shouldn't be in the second tier. On the other side, some users directly called out the desktop app: Weston said the Droid desktop app is terrible, and Hassan said it is hard to use compared with Codex desktop. Benchmarks and efficiency were praised, while the desktop experience was questioned.

The community is also filling in features themselves. Daniel White wants to use Claude sub in Droid. AI Jiahao ported a t3code-style Claude Code integration into Droid, said usage still looked healthy when running mission mode, and shared a repo link.

Droid 1.0 is available today. The desktop app supports macOS and Windows, and there is also a terminal UI and SDK.

Links:
Factory release post
Technical blog
Desktop download
terminal UI
Droid SDK
Community bridge

It won on the benchmark leaderboard, but the desktop app still has catching up to do. The comment section has already done the math on experience.