if you build with Claude Code, here's how to make it use fewer tokens with one new setting, and when it backfires.
the setting is Code Mode: Claude writes code that makes several tool calls in one step. turn it on with one environment variable:
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1
it's experimental, off by default (v2.1.323 or later) and only runs in the interactive terminal: not claude -p, the Agent SDK, the VS Code panel, the desktop app or cloud sessions.
my test (0:13): the same jobs with Code Mode off and on, three pairs of runs on Opus 5.5. tokens dropped 36 to 38% in every pair. the catch: my first run cost 83% more at api prices, because it wrote far more to the one-hour cache, and on Opus 5.5 that write costs 40 times a read ($8 vs $0.20 per million tokens). the other two pairs came out 15% and 17% cheaper. so check your cost, not just your tokens. Anthropic's page never mentions tokens, so the token numbers here are from my own test.
when to switch it on: Anthropic says tasks with many related tool calls, like a change that repeats across many files.
the practical (10:47): Gabriel Petersson's prompt, "make 10 different increasingly high variance versions and show me, then i can pick the best one". i ran it on the hook of my Threads post with 444K views. number 1 stayed close: "okay, why the hell isn't anyone talking about this?" number 10 went far: "i found out something today and honestly wish i hadn't." try it on designs, landing pages and plans too, then pick your favorite parts of each.
also this week:
- Qwen4-Max and Hy4 flash (2:45): Hy4 flash is $0.07 in / $0.28 out per million tokens, and Tencent says it's close to Haiku 5.5 in a blind test with 85 of its own users. Qwen4-Max is $1.50 / $4.50, and Alibaba says cost per task drops about 60% on its own test. Haiku 5.5 is $0.10 / $0.50 (the $0.10 is for prompts up to 100,000 tokens), so Qwen4-Max costs more than Haiku.
- Anthropic's guide to always-on agents (4:14): an event inbox, timers, memory, permissions and an outbox. the agent sleeps most of the time, your passwords sit in a proxy outside it, and memory only holds what it can't look up again. my version: one rules file it loads every run, plus one progress file per job.
- iPhone Duo (6:53): Claude Code on desktop can run your app on an iPhone Duo in the iOS simulator, then fold and unfold it to test each layout. beta, Xcode 27.1 or later.
- Hugging Face and OpenRouter (8:07): one token and one endpoint for every open model on the Hub that a provider serves, no markup from Hugging Face, setup guides for Claude Code and Codex, and $2 of credits a month on PRO.
- Jensen Huang (9:30): a source told CNBC Nvidia's board is not considering bringing him back, and a petition to reinstate him has 180,000 signatures (per CNBC). after filming, Nvidia reported $115.8B revenue for the quarter, up 103% on a year ago. that was Jensen's quarter.
today's sponsor is my software, Threadify: your lead generation agent for Threads, X and YouTube. it tracks your links, so you see which posts actually drive clicks, then you do more of what's working. free 7-day trial, card required:
https://www.threadify.app/home?utm_source=podcast&utm_medium=audio&utm_campaign=lab-0045&utm_content=sponsor&video_slug=lab-0045&cta_slot=podcast_description&entry_angle=whats_working
links:
- Code Mode docs: https://code.claude.com/docs/en/code-mode
- Anthropic prices (cache, Haiku 5.5): https://platform.claude.com/docs/en/about-claude/pricing
- Hy4 flash (Tencent): https://x.com/i/status/2122217770314301469
- Qwen4-Max (Alibaba): https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max
- always-on agents guide: https://claude.com/resources/articles/building-always-on-agents
- iPhone Duo simulator docs: https://code.claude.com/docs/en/desktop-ios-simulator
- Hugging Face and OpenRouter: https://huggingface.co/blog/openrouter
- CNBC on Jensen Huang: https://www.cnbc.com/2026/11/16/nvidia-ceo-jensen-huang.html
- Nvidia's results (SEC): https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000088/q3fy27pr.htm
- Gabriel Petersson's prompt: https://x.com/i/status/2122083139971416106
- my Jensen take (LAB 0042): https://youtu.be/JelRGipwZ_Y
0:00 make Claude Code use fewer tokens with one setting
0:13 Code Mode: my test, the catch, when to switch it on
2:10 this is the LAB
2:28 today's thread: the best model wins
2:45 Qwen4-Max and Hy4 flash vs Haiku 5.5
4:14 Anthropic's guide to always-on agents
6:12 today's sponsor: Threadify
6:53 iPhone Duo in the Claude Code simulator
8:07 Hugging Face and OpenRouter: one token for open models
9:30 Jensen Huang: the board, the petition, the earnings
10:47 the practical: 10 high-variance versions
13:26 the best model wins, so test it on your own work
watch the full episode: https://youtu.be/863KczMCX6w