---
title: "Caveman news"
description: "Releases, research and letters from the people building Caveman. Each piece is also Markdown at its URL plus .md."
canonical: https://caveman.so/news
last-updated: 2026-11-12
---

# Caveman news

Releases, research and letters from the people building Caveman. Each piece is
also Markdown at its URL plus `.md`.

- [Introducing Caveman Code](https://caveman.so/news/introducing-caveman-code) (Release, 2026-11-12): The most efficient coding agent, in one app with the agents you already use. On the same model it finished Terminal-Bench 4 for 21% less than Codex and 14% less than Claude Code.
- [Benchmarking Sonnet 5.5: less thinking for the same answers](https://caveman.so/news/benchmarking-sonnet-5-5) (Field report, 2026-09-28): Measured on our 48-task suite, not on your traffic. At the same effort level, Sonnet 5.5 solved at least as many tasks and usually spent less.
- [Caveman, RTK, Headroom, and context-mode compared](https://caveman.so/news/caveman-vs-rtk-headroom-context-mode) (Guide, 2026-09-19): What each tool shrinks, how you recover omitted content, telemetry defaults, and what the published numbers measure.
- [21 adapters for the agent you already built](https://caveman.so/news/21-adapters-for-your-agent) (Release, 2026-09-19): Eight TypeScript adapters, thirteen Python adapters, and recoverable tool-result compression inside the agent you already run.
- [How to cut Claude Code costs without making it dumber](https://caveman.so/news/how-to-cut-claude-code-costs) (Guide, 2026-09-07): Seven levers, the measured number behind each, what breaks them, and the cases where they lose money.
- [How Caveman works with LiteLLM](https://caveman.so/news/caveman-and-litellm) (Guide, 2026-09-07): Per-call api_base, fleet-wide config.yaml, Claude Code through both, and a routing callback that never touches inference.
- [Headroom vs Caveman for prompt compression, measured](https://caveman.so/news/headroom-vs-caveman) (Field report, 2026-09-07): Same six tasks, 54 runs, an exact-answer check on every one. 33.2 percent against 6.7, and the rows where each lost.
- [What Claude Code actually bills you for](https://caveman.so/news/what-claude-code-bills-you-for) (Guide, 2026-09-07): Four counters, one of them ten times cheaper than the others, and a prefix that decides which one you pay.
- [MCP servers cost tokens before you type anything](https://caveman.so/news/mcp-tool-overhead-in-claude-code) (Guide, 2026-09-07): 224,655 characters of tool schemas on one first turn. How to measure yours and three ways to bring it down.
- [Does the Caveman skill actually save money? Sometimes](https://caveman.so/news/does-the-caveman-skill-save-money) (Note, 2026-09-07): 65 percent fewer output tokens on ten prompts, 1,000 to 1,500 input tokens per turn to get it, and plans where it saves nothing.
- [Cutting token costs in Cursor, Codex CLI and Gemini CLI](https://caveman.so/news/cut-token-costs-in-cursor-codex-and-gemini-cli) (Guide, 2026-09-07): The same two levers as Claude Code, the exact install for each agent, and the meters where token savings do not help.
- [How to measure whether a token optimization saved anything](https://caveman.so/news/how-to-measure-agent-token-savings) (Guide, 2026-09-07): Paired tasks, provider counters, a correctness gate, the spread, and four labels that stop an estimate becoming a claim.
- [How removing tokens raised the bill](https://caveman.so/news/how-removing-tokens-raised-the-bill) (Field report, 2026-08-13): A token counter said 96.2 million saved. Paired trials found a higher bill. The denominator explains both.
- [Your agent pays before it works](https://caveman.so/news/your-agent-pays-before-it-works) (Field report, 2026-08-13): One real agent loaded 224,655 characters of tool schemas before its first action. Most came from plugins it might never call.
- [The 47 tokens your agent should never summarize](https://caveman.so/news/the-47-tokens-your-agent-should-never-summarize) (Field report, 2026-08-13): Compaction can preserve the job while deleting its rules. One benchmark found a 47-token buffer prevented the failure.
- [The average stayed flat. 125 tasks changed outcome.](https://caveman.so/news/the-average-stayed-flat) (Field report, 2026-08-13): A 998-task replay gained seven net passes while 125 outcomes flipped. Aggregate scores hid both directions.
- [A benchmark printed 26.5%. We refused the result](https://caveman.so/news/we-cut-26-percent-and-refused-the-claim) (Field report, 2026-08-13): Nine quality-held pairs produced an attractive diagnostic and three publication blockers. The number stayed out of our claims.
- [Caveman is live on GreenPT](https://caveman.so/news/caveman-live-on-greenpt) (Release, 2026-08-03): The skill now runs at the API level of a green AI platform. No install, nothing to maintain. First platform integration.
- [Efficiency and the token economy](https://caveman.so/news/efficiency-and-the-token-economy) (A letter, 2026-07-25): The price of a token collapsed and the bill went up. What follows from reading that correctly.
