The Information Machine
FollowingDay 2in the edition since 17 Nov 20262 sourcesupdated 17 Nov 2026

Token costs in AI coding agents

The gist

Latest: UC Berkeley and Microsoft's TokenEconomizer cuts coding-agent costs 18-31%

Token costs in multi-turn coding agents compound across sessions because each turn carries the full conversation history. A drop-in external controller that cuts those costs without touching the agent or degrading task success rates is a practical reduction available to any team running these agents.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

UC Berkeley and Microsoft built TokenEconomizer, an external controller that reduces coding-agent token costs by 18.4% to 30.9% by learning from past sessions. Each agent turn resends the full growing conversation history, compounding costs across a session. TokenEconomizer sits outside the agent as a black box, modifying only prompts and tool settings without altering the agent itself. It learns which facts to inject, which tools go unused, and when to abort a failing loop. Tested on Claude Code, Codex, and GitHub Copilot CLI across 277 SWE-Bench tasks, it improved Pass@1 in 4 of 5 agent-model configurations and held savings across every repository tested. The plugin is open-source and available on GitHub.

How it developed
First in the edition · 17 November 2026
16 November 2026in the 17 Nov edition

Three tools and a study published within a week address coding agents resending full conversation histories each turn.

TokenEconomizer, from UC Berkeley and Microsoft Research, cut mean per-task cost 18.4% to 30.9% on 277 SWE-Bench tasks. ByteAsk reduced pooled tokens from 1,795K to 395K per run for 68.1% savings on C/C++. Reason Machines reported 22% lower cost than Codex via prompt-prefix caching. A Carnegie Mellon and JetBrains study found token spending tracks AI usage, not shipping.

13 November 2026

Deepnote launches Agent Initiative Benchmark for proactive agent decision-making; GPT-6 Astra scores highest, Sol best on cost-performance

11 November 2026

NaceAI releases Drex document processing system, scoring 0.834 on Parse Index at $7 per 1,000 pages

10 November 2026

Route Anything benchmark from UC Berkeley evaluates 452 models across 25 tasks, finding 8.82x cost efficiency via learned routing

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy