ZeroNoise Logo zeronoise
Post
Claude Code's experimental Code Mode, and the real cost of Gemini 4 reasoning effort
•
6 min read
• 98 docs
Claude Code can now chain tool calls through generated JavaScript. Meanwhile Gemini 4 users are settling which variant and effort level to use, Antigravity's harness is claiming coding-benchmark gains, and practitioners shared a small-eval method for switching to cheaper models.

Claude Code Code Mode: scripts instead of one tool call at a time

Claude Code has an experimental mode where Claude writes and runs JavaScript to chain tool calls. Normally it calls Bash, Read or Edit one at a time and waits for each result. In Code Mode a script can run commands in parallel, pass arguments between them, and filter outputs before they reach the context window . Addy Osmani gives two cases where this helps: large search results that would otherwise go straight into context, and the same refactor applied across dozens of files .

How to try it:

  • Start a session with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude, or set the variable in ~/.claude/settings.json .
  • In auto or bypassPermissions mode, scripts such as test suites or builds can run in the background for up to two hours while you keep working. /tasks lets you monitor and control them .
  • Press Ctrl+O to see the JavaScript Claude generated .
  • Guardrails still apply. Every generated tool call goes through your deny and ask policies, sandbox and PreToolUse hooks. You can attach hooks to the REPL tool to inspect scripts before they run or to redact tokens from output .

Suggested first uses are multi-file audits, batch edits and parallel search-and-filter jobs. Send problems via /feedback .

Gemini 4: which variant, which effort level

Logan Kilpatrick's quick guide: Carbon has "big model smell," Neon is "best all around," and Hydrogen is "quick and efficient" . He says there's no reason to use Argon now, since it was only the first Gemini 4 revision .

Effort level makes a big difference in cost. In Simon Willison's pelican test, Hydrogen on high used 5,311 output tokens, cost 3.99¢ and took 19s. On xhigh it used 31,477 tokens, cost 23.61¢ and took 2m36s, for a noticeably richer image . That's roughly 6x the cost and 8x the time. His comparison grid for Carbon, Neon and Hydrogen at every reasoning level is here .

Opinions on Neon still differ. Perplexity's Aravind Srinivas runs Neon on High in Perplexity Computer and has moved some teammates to it . Jeremy Howard says DeepMind described it at launch as only "approaching" the frontier, and that evals are "a very imperfect measure of model utility" . Nicholas Moy disagrees and says many people knew how good it was . On the other side, one user says Neon makes mistakes when trying to prove bugs exist, and would pick Sol, Astra or Opus for everything they tried .

Antigravity: harness claims and more quota

Google's Varun Mohan says Neon running in the Antigravity CLI is #1 on the AA Coding Agent Index. He also reports a large jump over mini-swe on SWE-sweep, a bug-finding and fixing eval on large codebases, and says the gain is bigger than other frontier models get from their own native harnesses. More harness improvements are due within a week . These are Google's own numbers. Moy says his preferred setup is remote agents over the web, and he "barely open[s] an actual IDE anymore" .

Mohan also reset everyone's Gemini quota on Antigravity. He says caching improvements mean a quota now covers 22% more work . That speaks to last brief's complaint that Neon was using up limits too fast because of low cache hit rates.

T3 Code: bring your own subscriptions

Theo says T3 Code has officially supported Antigravity through Google's integration for months, and there is "zero risk" in using Gemini subscriptions with it . DeepMind's Patrick Toulme, who has landed PRs improving T3 Code's Gemini support, says users "will NOT be banned ever" for using Antigravity in T3 Code . According to Theo, 99% of T3 Code usage runs on subscriptions rather than API pricing . A new feature lets you sign in to MCP servers on remote machines without going through the terminal .

Use your own small evals to pick cheaper models

On Riley Brown's channel, Suhail Doshi laid out a simple method. Make a folder called eval with five or six tasks you do often. Paste the same prompt into each model and judge the output. Doshi argues this can reveal something like a 30x price cut for the same result . On his own design tasks he saw almost no difference between GPT Astra and DeepSeek V4 4.1 Flash, at about one-thirtieth the cost . For something more automated, take the trace of a task you got right and ask a frontier model to turn it into a scored eval you can rerun whenever a new model comes out . He suggests OpenRouter for switching between models quickly .

Riley also says skipping the GUI works better. Codex edits video better "fully headless" with ffmpeg than through DaVinci Resolve's MCP . Doshi thinks models have simply seen far more ffmpeg code . Riley says the same holds for making PDFs directly instead of through Canva, and for Blender, where every action maps to Python .

Before routing sensitive code through an aggregator, note the OpenRouter report circulating this period. A researcher sent secret links through the service, and someone opened them 40–130 hours later. The cheaper providers AtlasCloud, Novita and StreamLake were apparently forwarding requests to MiniMax . Sebastian Raschka's explanation: when cheaper providers hit capacity, they reroute traffic to the official model provider rather than return an error . Who opened the links has not been explained.

LangChain's "knowledge factory" for agent context

LangChain describes how its internal agents stay supplied with knowledge . It sorts that knowledge into three kinds: skills (how to do the work), domain knowledge (how the business works) and data-access guidance (where current data lives) . Working agents get a tool to flag knowledge that is missing or wrong. A separate curator agent runs on a schedule, reads those flags along with traces and raw sources, and updates the knowledge store. The agents doing the work don't edit it themselves, which LangChain says would fill it with contradictions . The advice is to start with one recurring workflow and close that loop rather than documenting everything first .

Smaller notes

  • Codex worktrees: swyx found 20GB of duplicated node_modules across worktrees . He says that for about a month Codex hasn't been able to clean up its own worktrees, and he has given up .
  • Codex desktop: the latest ChatGPT/Codex app update lists "more reliable Codex Remote" .
  • Kent C. Dodds uses Jev in Kody to re-check vector-search results. When Jev judges a hit very likely to be exactly what the agent wants, Kody expands its details .
  • Game dev: DesignArena ranks GPT-6.2 Astra (high) #1 and Sol (high) #3 in its Game Dev arena. Astra averages $0.98 and 6.8 minutes per generation, Sol $0.22 and 4.6 minutes .
  • Multi-bot setup (from a sponsor segment): Matthew Berman says Lauren Tan runs a team of Grok-powered coding bots that coordinate in Slack and hand work off to Cursor, reaching 2,500 PRs in a month. Each bot has its own cloud machine .
Claude Code's experimental Code Mode, and the real cost of Gemini 4 reasoning effort
Simon Willison

In a Gemini 4 Hydrogen pelican-and-bicycle output comparison, xhigh produced a richer result than high, but used 31,477 versus 5,311 output tokens and cost 23.61¢ versus 3.99¢; reported runtimes were 2m36s versus 19s. This is a substantial quality-versus-cost-and-latency trade-off when choosing reasoning effort. Willison also shared a comparison grid for Gemini 4 Carbon, Neon, and Hydrogen across reasoning levels.

Gemini 4 tip: the difference between Gemini 4 Hydrogen on "high" thinking effort and that same model on "xhigh" is VERY significant Here'… Here's the full comparison grid for Gemini 4 Carbon, Neon and Hydrogen across all of their reasoning levels [https://static.simonwillison…
Riley Brown
Profile
  • To choose a cheaper model, keep an eval folder of recurring tasks, run the same prompts across models, and compare output quality against cost; even five or six examples can provide a useful barometer. For a more automated eval, save a successful task trace and ask a frontier model to turn it into a scored test. OpenRouter is one suggested place to quickly switch between models and try them.
  • For video editing, Riley found that asking Codex to manipulate files directly with ffmpeg worked better than using DaVinci Resolve through MCP; he attributes the difference to models’ greater exposure to ffmpeg code. He also describes Blender workflows through Python, with Cycles rendering and Runpod GPUs for work too slow to render locally.
  • For rapid multiplayer-game iteration, Riley used Opus 5.5 with a Convex real-time database hosted on Vercel: he collected chat suggestions, pasted them as one prompt, then played the updated version live.
Is Claude Still Worth Paying 30x For? | Billion-Dollar Founder Suhail Doshi
swyx

A post alleged that someone accessed outputs sent to MiniMax after the author sent secret links that were opened 40–130 hours later; it also alleged that AltasCloud, Novita, and StreamLake forwarded requests to MiniMax. A reply said OpenRouter usually selects the cheapest inference provider, but when cheaper providers are at capacity it may reroute traffic to the model’s official provider instead of returning an error. For coding-agent workflows, this is a reason to check provider routing and data-handling practices before sending sensitive code or prompts through an inference aggregator.

the OpenRouter situation for anyone curious: -someone managed to access outputs sent to MiniMax -researcher discovered this by sending a … [@swyx](https://x.com/swyx) OpenRouter usually uses the cheapest inference provider. The (cheaper) inference providers are sometimes at c…
Riley Brown
  • Riley described a live game-building loop using Opus 5.5, Convex for real-time multiplayer, and Vercel hosting: after playing for about 20 minutes, he collected chat suggestions, pasted them into one prompt, then played the updated version.
  • For video editing, Riley said Codex worked better when asked to edit files headlessly with FFmpeg than when using DaVinci Resolve’s MCP integration; Su suggested this may reflect greater model exposure to FFmpeg code than to computer-use interactions or that integration.
  • To compare models on recurring work, make a small eval folder, run the same prompts across models on roughly five or six examples, and compare quality with cost; a successful task trace can also be turned into a scored eval with help from a frontier model. Su recommended OpenRouter for trying models, and named DeepSeek and GLM as preferred open-weight options, with their Flash versions cheaper and a 3-trillion-parameter Kimi model suited to some more frontier workloads.
  • Particle’s founder said its Particle One model, post-trained from DeepSeek Flash 1017 with reinforcement learning, was 42% more token-efficient and completed tasks correspondingly faster; he described targeting eval data that exposed inefficient token use, while noting that customer results varied by task category.
Is Claude Still Worth Paying 30x For? | Billion-Dollar Founder Suhail Doshi
Matthew Berman

In a sponsored segment, the video says SpaceX engineer Lauren Tan (“Potato”) uses a team of coding bots that coordinate in Slack, hand work to Cursor, and keep iterating; it credits Grockbot with powering/driving Cursor and reports 2,500 PRs in one month. Each bot gets its own cloud computer, where it can work on an assigned task while the user’s laptop is shut.

I’m Crashing Out About Anthropic
Riley Brown

Riley Brown’s conversation with Suhail covers building custom evaluations for real or non-coding work and deciding when to switch to cheaper open models; its chapter list also flags AI pricing, subsidies, and always-on agents, but gives no implementation steps or switching criteria.

Is Software Engineering Dead? Maybe. What should you do about it? "Scale up your ambition" according to [@Suhail](https://x.com/Suhail) I…
Addy Osmani

Claude Code’s experimental Code Mode writes and runs JavaScript to chain tool calls, parallelize commands, and filter results before they enter context; this can make multi-file audits, batch edits, and refactors across dozens of files more efficient. Enable it with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 when starting a session or in ~/.claude/settings.json. In auto or bypassPermissions mode, scripts can run in the background for up to two hours while you keep working in the session; use /tasks to monitor and control them. Generated tool calls still follow existing permission policies, sandbox constraints, and PreToolUse hooks.

Code Mode is an experimental new feature in Claude Code that can improve both speed and context efficiency when Claude needs to orchestra…
swyx

swyx says worktrees consume 20 GB through repeated node_modules directories . In a related post, he reports that Codex had been unable to clean up its worktrees for about a month despite being asked, and that he gave up .

worktrees must die this is 20GB of repeated node\_modules lol ![](https://pbs.twimg.com/media/HPZhNpXbsAAn_-G.jpg) something broke in the last. month or so and codex is unable to clear its own worktree hell despite me asking nicely. giving up ![](https…
Logan Kilpatrick

A terse model-selection shorthand associates Carbon with “big model smell,” calls Neon “best all around,” and describes Hydrogen as a “quick and efficient model”; the post does not identify the referenced models or give selection criteria.

[@SahilPanhotra](https://x.com/SahilPanhotra) Carbon = big model smell Neon = best all around Hydrogen = quick and efficient model
Kent C. Dodds 🐨

Kent says that once Kody.codes clicks, tasks start to feel worth automating . In a recent example, he used it to ask “Howie · Home Manager” for an age-appropriate playlist for his 12-year-old son and friends; it created 60 clean songs on the Dodds Kids Spotify account and saved the playlist as a Sonos favorite .

When [https://kody.codes](https://kody.codes) finally clicks for you, everything becomes automateable and "worth the effort of automation… [@thorstenball](https://x.com/thorstenball) This is one of the first things I used [https://kody.codes](https://kody.codes) for actually.…
Kent C. Dodds 🐨

Kent C. Dodds uses Jev in Kody after vector search: when Jev judges a result very likely to be exactly what the agent is seeking, Kody expands that result’s details; he also uses Jev for many professional and personal automations. A quoted user, @asmartbear, reports that Jev was difficult to get to match Sonnet on many classification jobs.

I'm using Jev in Kody to improve search results. If a search result comes back from the vector search and Jev determines it is very highl… Is anyone actually using Jev? 1mo ago the hype was amazing; if you search on YouTube you find videos from then, but none now. I tried it …
ThePrimeagen

ThePrimeagen strongly endorses using tree visualizations in prompts, calling them “the best thing ever.”

TJ is absolutely correct Tree visualization on prompts are in fact the best thing ever
Jason Zhou

An open-source job-hunting agent, Log Apply, uses Treg’s jobs API, AsideAI, Claude Code, and Log to find roles, create a job-specific resume and cover letter, fill out the application, and apply while the user is at work. Jason Zhou described it as an open-source job-hunter agent built on Treg.

introducing log apply your agent applies to jobs for you while you're at work it finds the role writes a resume and cover letter for that… OpenSource Job hunter agent built on top of [@treg_ai](https://x.com/treg_ai) [https://x.com/deepman_1/status/2123447650037158195](https:…
Logan Kilpatrick

For Gemini 4, Logan recommends Carbon or Neon over Argon, describing Argon as only the first Gemini 4 revision.

[@RichmanRonald](https://x.com/RichmanRonald) [@PatrickToulme](https://x.com/PatrickToulme) [@SahilPanhotra](https://x.com/SahilPanhotra)…
Theo - t3.gg

A post about Antigravity paired a reset of everyone’s Gemini quota with the claim that caching improvements increased the work users could get done per quota by 22%.

Reset everyone’s Gemini quota on Antigravity. The caching improvements have increased the amount you can get done in quota by 22%. Enjoy …
Greg Brockman

DesignArena’s Game Dev arena ranks GPT-6.2 Astra (high) #1 and GPT-6.2 Sol (high) #3. DesignArena reports average generation times and costs of 6.8 minutes and $0.98 for Astra, and 4.6 minutes and $0.22 for Sol; it places both on its preference-versus-price and preference-versus-speed frontiers, and says Sol delivers similar output quality to GPT-6 Astra (max) at about 4× lower cost and 40% faster. Brockman says models that excel at game development could lower the barrier to making games for anyone with an idea.

BREAKING: GPT-6.2 Astra (high) and Sol (high) by [@OpenAI](https://x.com/OpenAI) are ranked [#1](https://x.com/hashtag/1) and [#3](https:… I played computer games growing up, and remember wondering how I could learn to make my own games. Now we’re creating models which are ex…
LangChain Blog

LangChain says its internal GTM and content-studio agents use a knowledge factory, and LangSmith Engine maintains one per agent. For an agent workflow, keep reusable knowledge in three buckets: procedural skills, durable domain knowledge, and guidance on where and how to access current data. Have agents flag missing or incorrect knowledge, then use a separate curator agent on a cadence to review those flags alongside raw sources and agent traces and update the shared knowledge; skills can be stored in GitHub. Start with one recurring workflow, inspect where agents struggle, and close the loop rather than trying to document everything upfront.

How to build a knowledge factory
Nicholas Moy

Arav Srinivas says Gemini Neon with Reasoning set to High feels both intelligent and fast; he is using it in Perplexity Computer and has moved some teammates to the model.

It’s been a while since I felt a model was intelligent and fast at the same time and is too good to be true. Gemini Neon (Reasoning set t…
Theo - t3.gg
  • Theo identified the developer who posted about T3 Code as a Google employee; the developer said they now use Neon exclusively in T3 Code and had shipped their first PR to improve the tool because they liked it.
  • T3 Code users can bring existing subscriptions rather than relying only on API pricing; Theo said 99% of usage is subscription-based, including Gemini.
Google employee btw [https://x.com/thecsguy/status/2124213504827646111](https://x.com/thecsguy/status/2124213504827646111) i use neon exclusively in t3code now today i shipped my first PR to t3code. when a tool is this good you want to help make it even better! 🔥 [@brotchie](https://x.com/brotchie) You can absolutely bring your subs. 99% of T3 Code usage is with subscriptions (including gemini!) I'…
Greg Brockman

Greg Brockman highlighted OpenAI’s top ranking in Arena’s updated Alignment Index: GPT-6.2 Sol ranked #1, and OpenAI held all seven top positions. The index measures agent behavior in real user sessions across unauthorized actions, deceptive completion, and false attribution—useful reliability context, but not evidence of coding performance.

OpenAI models topping the alignment benchmarks: [https://x.com/arena/status/2123835034532729205](https://x.com/arena/status/2123835034532… The Arena Alignment Index has been updated! We've refined how our signals detect model failures, providing broader coverage and more reli…