--- title: "Caveman Code" description: "The most efficient coding agent. On the same model, it finished Terminal-Bench 4 for 21% less than Codex, and day-long work across connected apps for 76% less than Claude Code. It did that on 31% fewer tokens than Codex per solved task, and 83% fewer than Claude Code per session." canonical: https://caveman.so/code last-updated: 2026-11-12 --- # Caveman Code The most efficient coding agent. On the same model, it finished Terminal-Bench 4 for 21% less than Codex, and day-long work across connected apps for 76% less than Claude Code. It did that on 31% fewer tokens than Codex per solved task, and 83% fewer than Claude Code per session. | | | | --- | --- | | For | Developers and their teams | | Runs | On your Mac, in your browser or on a Cloud computer; it keeps working while you sleep | | Built on | Our own harness, under our own agent. Claude Agent, Codex and Cursor run beside it on your own logins | | Get it | [Sign in](https://app.caveman.so/signup), or [download the Mac app](https://caveman.so/download). Free on your own keys and logins | ## Our agent runs on the most efficient harness Same model, same work, 21% to 76% less. A harness is everything around the model: the prompt, the tools and the loop that runs them. Caveman, the agent built into Caveman Code, runs on ours. We ran it against Claude Code and Codex, each on its own lab's model: the model is the same, so the difference in the bill is the harness. | Against | Model | Work | Their cost | Caveman Code | Change | Tokens | Result | | --- | --- | --- | --- | --- | --- | --- | --- | | Claude Code | Claude Opus 5.5 | Terminal-Bench 4: 45 coding and system tasks in a terminal, 4 runs each, at high reasoning | $6.04 a solved task | $5.19 a solved task | 14% less | 22% fewer | 97 of 180 runs solved, to Claude Code's 83 of 177 | | Claude Code | Claude Opus 5.5 | Three working days: three day-long sessions, each one conversation of 15 to 17 requests across four connected apps, run once | $7.65 a session | $1.83 a session | 76% less | 83% fewer (3.7M to 21.6M a session) | 48 of 48 steps passed, to Claude Code's 46 | | Claude Code | Claude Opus 5.5 | Connector tasks: 24 tasks that reach other apps through MCP servers, 2 runs each | $0.431 a task | $0.127 a task | 71% less | 81% fewer (74k to 383k a task) | 48 of 48 runs passed, to Claude Code's 46 | | Codex | GPT-6.1 Sol | Terminal-Bench 4: 48 coding and system tasks in a terminal, 4 runs each, at high reasoning | $1.52 a solved task | $1.21 a solved task | 21% less | 31% fewer (3.2M to 4.7M a solved task) | 90 of 192 runs solved, both | | Codex | GPT-6.1 Sol | Connector tasks: 24 tasks that reach other apps through MCP servers, 2 runs each | $0.109 a task | $0.069 a task | 37% less | 49% fewer (88k to 173k a task) | 48 of 48 runs passed, both; cheaper on 24 of 24 tasks | Each pair ran the same model on the same work, the lab's harness against ours with no router. Cost is list price, metered at one proxy. Claude Code 2.1.309 on Terminal-Bench 4, 2.1.318 on connector tasks, 2.1.319 on the working days; Codex 0.163 on Terminal-Bench 4, 0.165 elsewhere. Our own runs, November 2026. - These runs measure the harness in its benchmark build. The app ships the same harness with three benchmark-only settings turned off, and only Caveman, our own agent, runs on it: a Claude plan runs as Claude Agent, on Anthropic's harness. - Tokens are every token the model processed: input, cached or not, and output, counted at the same proxy as the cost and over the same runs. Per solved task on Terminal-Bench 4, per task and per session elsewhere. - Terminal-Bench 4 against Claude Code: 45 tasks, 4 runs per task per harness, in the same hours. We solved 97 of 180 runs to its 83 of 177 and cost 14% less per solved task ($5.19 to $6.04). The rig's 50-minute limit was lifted to two hours on the nine tasks where it cut either side off. - The largest savings against Claude Code are on connector work through MCP: 24 tasks run twice, and three working-day sessions run once each. Both suites pass almost everywhere, so they measure cost. Claude Code's missed steps come from a task whose frozen date reads as past, and one recall check. - The Terminal-Bench runs used the harness as it stood on November 2. The connector runs used the next version, which reaches connected apps through a program so their raw results stay out of the conversation. That change is most of the saving on connector work, and it is the version in the app. - Terminal-Bench 4: the 48 of 66 tasks that run in E2B cloud sandboxes, 4 runs per task per harness, interleaved in the same hours. Cost per solved task counts failed runs too. 16 of the 48 tasks were held out from all tuning; on those both solved 26 of 64 and ours cost 19% less. - On DeepSWE's 53 held-out tasks, one run each, Codex solved 33 to our 30, a gap inside the run's 90% interval, and ours cost 17% less per solved task. ## Every agent you use, in one window, on what you already pay for Caveman Code is a Mac app that also runs in your browser. Claude Agent, Codex, Cursor and Caveman, our own agent, work side by side in tabs and splits. - **Bring what you already pay for.** Your Claude plan runs as Claude Agent and your ChatGPT plan as Codex, each on its own login. Caveman, our agent, takes any API key, coding plan or gateway you add. Free on your own keys and logins. - **A Chief runs your chats.** Tell it what you want done. It starts the chats, answers their questions, reports back, and goes ahead on its own when you don't answer in time. - **A guide to every pull request.** Written before you open it: a one-line TL;DR, the one call that is yours, then each change as before, after and why. - **Agents that use a computer.** On a Cloud computer an agent drives a Linux desktop and a browser to check its own work. Watch it live, take over, hand back. On your Mac it works in your apps, in the background. - **Close the lid. It keeps going.** Turn it on, and working chats continue on a Cloud computer when the lid closes. Add your other Mac with a one-time code and move a chat there too. Cloud computers are in beta. - **Also in the app.** Automations that run by hand, on a schedule or on a GitHub event. A worktree for every chat. Agents that start agents and message each other. Plugins from Claude Code, Codex and Cursor. One memory that every agent reads. A terminal app is coming soon. ## Review built for the agentic era Agents open more pull requests than anyone can read. Caveman reads each one first and writes you a guide, ahead of time, on the cheapest model that can. - **TL;DR first.** Every pull request opens on one line that says what changed. - **Before, after, why.** Each stop in the guide shows the change and the reason for it. - **Your call.** The one judgment only you can make is marked as yours. - **Mark read, move on.** One key per stop. Select any line to ask, fix or comment, and an agent picks it up. Caveman Code is not open source. Already on Claude Code or Codex? The open-source [Caveman CLI](https://caveman.so/caveman) makes that agent cheaper.