Halv halves your coding agent's tokens — and solves more tasks
On 20 SWE-rebench tasks: 7→10 solved at half the tokens per fix, $10/mo. Plus NightRun boots an LLM off USB, and Copilot retires 4 models Oct 2.
Copy markdown
7→10 tasks solved, half the tokens per fix
On 20 paired SWE-rebench tasks driving Codex, Halv lifted the solve rate from 7/20 to 10/20 while cutting tokens per correct answer from 9.06M to 4.43M — a 51.1% drop, counted across all runs including failures. Cheaper agents that also finish more work.
Where the tokens were going
Halv's engine is three parts: Headroom strips duplicated files, stale history and repeated instructions from context; Crux is a code index that answers "who calls this?" without the agent re-reading the repo; RTK filters test logs, diffs and dependency trees before they hit the model. Most agent spend is re-discovery, not reasoning.
$10/mo, over subscriptions you already pay for
It's a desktop app (macOS, Windows, Linux) that runs Claude Code, Codex, Kimi and GLM side by side through your existing logins — no extra API keys. $2/mo with capped savings or $10/mo unlimited, 7-day trial, no card. Caveat: closed-source, and the benchmark is the vendor's own 20-task run.
Elsewhere: NightRun boots an LLM off a USB stick
A new MIT-licensed, UEFI-resident runtime in Rust boots straight from USB or SD card into RAM and runs Llama 3.2, Granite 4.1 or Qwen3 4B fully offline — no OS, no network after load. ~3 tokens/sec generation on a Raspberry Pi 5. A weekend air-gapped AI appliance.
Elsewhere: Copilot retires 4 models Oct 2
GitHub Copilot drops Claude Opus 4.7, Gemini 3.5 Flash, Gemini 3.6 Flash and Kimi K2.7 Code on October 2. If any of your Copilot configs pin those, repoint them now before requests start failing.