ZeroNoise Logo zeronoise
Post
Gemini 4 Neon reaches Cursor and Antigravity; Grok Bot's harness lead explains how to cut agent cost
•
5 min read
• 207 docs
Gemini 4 Neon is now available in Cursor and Antigravity, and early reactions sharpen the debate over GPT-6.2 Astra. Also covered: Jediah Katz's cache and context tradeoffs for agent harnesses, LangSmith's measure of cost per merged PR, and advice on routing tasks across models.

Gemini 4 Neon arrives, and model-splitting becomes the default pattern

Google's Gemini 4 Neon is now available in the main coding harnesses. Cursor reports that at default effort it scores 63.7% on CursorBench 4.1, up from 57.3% for Gemini 4 Hydrogen, and calls that a new cost-performance frontier. Pricing is $3/M input and $15/M output tokens . Antigravity is rolling it out across the desktop app, CLI and IDE extensions for paid AI Pro and Ultra subscribers, and says it ranks #1 on the Artificial Analysis Coding Agent Index. Antigravity's Orchestrator can pair Neon as planner with Hydrogen as executor, so Neon's judgment goes "where it matters and Hydrogen's speed everywhere else" . One Google launch post claims Neon is cheaper, faster and "better behaved" than Carbon, and beats Opus 5.5 and Sol 6.2 on most benchmarks . Those are vendor claims.

Theo started out skeptical. By the end of the day he called Neon a "weirdly perfect counter balance to GPT-6.2 Astra… so much more pleasant to work with" . His current T3 Code setup gives each model a role, each running in its own harness on its own subscription :

  • Fable leads implementation
  • GPT-6.2 Astra playtests
  • Gemini 4 Neon writes all the copy

He explains the split with a distinction worth borrowing: capability ("at its best, what can it do") versus coherency ("does it understand what it's doing"). He says model families "have never felt so different" .

The Astra debate gets a concrete test case

Theo retested Astra after Tibo (OpenAI) said early testers had missed recent changes. His test question: why is creating a worktree slow on one VPS when two others are fast? The cause was a slow virtual disk, where the other machines had physical NVMe . Opus opened with "Slow disk, made worse by VDO." After the fixes, Astra was more concise but still led with second-order details . His conclusion: Astra "doesn't miss anything" but misses "the actual point" .

The takeaway for anyone evaluating debugging agents is to check whether the agent puts the dominant cause first, not just whether it is thorough. Theo says the system-prompt changes "helped a lot," but Astra still loses coherency after a few hours on long tasks and the improvements don't change his view .

Mathematician @nasqret, an early tester, has a different and more positive workflow:

  • He runs Astra in CLI agent mode with "no YOLO mode," watching every instruction it runs.
  • He starts from a very specific task driven by his custom "labyrinth" skill, then lets it run 1–4 hour research cycles with "Carry on" .
  • Astra won't set up multiple agents on its own, but will if asked, and can create communication channels between them .
  • He uses Lean to formalize the output. Astra points him to the hard proof steps to check .
  • He uses Sol as the faster, cheaper everyday agent .

Browser Use separately reports GPT-6.2 Sol as its top browser-use model: 82% (vs 81% for 6.1 Sol), 35% cheaper, 12 rather than 22 minutes, and 36% fewer steps .

Harness efficiency: what Grok Bot learned

Jediah Katz worked on Cursor's harness from early 2025 and now works on Grok Bot's. He says Grok Bot is 3x faster to reply and 2x more token-efficient since launch . The levers he describes apply to your own agent setups as well:

  • Keep the prompt prefix stable. Put anything volatile as late in the context as possible. A timestamp near the top causes a full cache miss .
  • Longer cache TTL gives more cache hits but costs more per token . In A/B tests, it paid off for Grok Bot's async coordinator, whose cache expired while it waited idle on subagents. It did not pay off for subagents that work continuously .
  • Every lever has a tradeoff. Compacting earlier, or moving tools into skills or a tool search, cuts context but adds steps. Cheaper models stumble more. Parallel tool calls and subagents cut steps but use more tokens .
  • Hold quality to a bar. Profile first, A/B test cost and speed, and confirm quality still meets the bar you set .

Measuring agent spend by merged PR

LangSmith for Coding Agents (public beta on LangSmith Cloud) attributes trace spend to pull requests, including threads that open several PRs and PRs that take several threads. It supports Claude Code, Codex, Cursor, Copilot in VS Code and Deep Agents . To set it up, install the plugin for each agent, connect GitHub and open the coding-agent overview page .

LangChain's usage advice applies with or without the tool:

  • Start fresh threads between tasks. Each new turn carries the whole conversation context .
  • Judge model changes by cost per merged PR, including moving subagents to smaller models .
  • Check skills that keep appearing in long, expensive threads. Their instructions may be causing extra turns .
  • Don't count all non-PR spend as waste. Review and investigation often don't end in a PR .

Routing and review across models

At Interrupt NYC, OpenRouter's Alex Atallah sorted tasks by entropy:

  • Low: for example, choosing who should review a PR. No frontier model needed .
  • Medium: PR review. Mostly open-weight models, with occasional frontier-model checks .
  • High: building new features. Use a frontier model .

He says writing with Opus and checking with GPT-5 noticeably improved code quality, while having Opus check its own work was "much, much worse" .

Harrison Chase recommends evaluating long-running agents on their whole trajectory and final environment state, not just the final answer. He suggests Harbor for sandboxed container runs . Turn unexpected failures into regression tests and rerun them whenever you change the prompt, model, context or harness .

Practitioner notes

  • Simon Willison on choosing models: pick a good default model family and ignore roughly 3% gains. Watch instead for new capabilities, such as 3D/Blender work . He once asked Opus 5.5 to make a game fun; it worked for 12 hours in Claude Code, ran 1,000 simulated matches, and made it fun . With the easy work delegated, he says every decision left to him is a hard one, and he has "never worked harder" .
  • Rewrites go stale. Daniel Stenberg says that while agents converted 185,000 lines of curl's C into 340,000 lines of Rust, curl merged 200+ bug fixes and some vulnerability fixes that the rewrites don't have .
  • Limit what unattended agents can do. Armin Ronacher's overnight AirPlay reverse-engineering agent decided at 3 AM to verify its work by turning a living-room speaker to max volume and listening through the laptop mic .
Gemini 4 Neon reaches Cursor and Antigravity; Grok Bot's harness lead explains how to cut agent cost
Riley Brown
Profile
  • Brown pairs Claude Code’s editable document and design artifacts, prompting @Claude to draft a newsletter in his voice and generate same-style images for each artifact type. He says Opus 5.5 is good enough for high-quality graphics and revises individual images by commenting directly on them. At week’s end, he gives Claude his phone notes and Twitter bookmarks to create a styled weekly-updates presentation.
  • For app-building, Brown prototypes an internal tool as an artifact for team feedback, then asks Claude Code to make it an app using GitHub, Convex, and Vercel; he says this setup creates a repository, database, and hosted app. His Final Draft-inspired script editor reportedly took three prompts plus two changes, and he had Claude search online for the app’s keyboard shortcuts.
  • In a sponsored Revolut Business example, he gave Claude Code the service’s API docs to build an MCP server for balances, transactions, virtual cards, team, and payment drafts. The agent matched invoices to saved recipients and prepared editor payments in pounds, euros, and Australian dollars after checking exchange rates; Brown retained manual review and sending.
Claude’s NEW Artifacts Are a Superpower (How I Use Them)
Salvatore Sanfilippo
Profile
  • For Redis memory-efficiency experiments, he is exploring compressed data structures and algorithms that trade higher CPU cost for lower memory use, then optimizing around cache locality to recover some overhead. In direct Astra–Fable 5.2 comparisons, he says Astra found significant corner-case bugs in Fable's implementations, while Astra often made further progress where Fable could not improve on Astra's solution.
  • He says Astra 6.2 did much of the latest work on UEC; he found it followed instructions well and produced fairly robust code, though it still had bugs, including in complex video-filtering work. For UI changes, he uses Opus for UX/UI advice and Fable as a second-model check on whether the changes are acceptable.
  • For sensitive security work, he says he used Mitos and describes Mos as an untiring collaborator that can fix critical bugs, assess exposure windows, and check continuously.
GPT 6.2 Astra dopo settimane di test
Harrison Chase
Profile
  • Chase recommends treating pull requests as events in an always-on coding-agent workflow: agents can work in the background in response to PR changes and comments, rather than only handling one-off coding requests.
  • Evaluate coding agents beyond their final answer: inspect traces and tool-call trajectories, actions and state changes, and test them in a sandboxed container. Chase names the open-source Harbor framework for sandboxed agent runs and recommends checking the code and final state afterward.
  • Turn unexpected agent failures into regression tests: sample traces, review them against a rubric with people or LLM judges, and rerun the cases after changing prompts, models, context, or the harness.
  • Set coding-agent objectives carefully: an instruction to “remove all bugs” could be satisfied by deleting the codebase, and agents pursuing goals persistently may break out of their sandbox.
Harrison Chase: what changes from 1 agent to 10,000
Simon Willison
Profile
  • Simon says agents let him produce in an hour software that previously took a week, raising the ambition of what he tackles . But delegating easy work leaves him with difficult, high-level decisions and a heavier, more tiring workload .
  • For a game project, he asked Opus 5.5 to make it fun; the agent worked for 12 hours and ran 1,000 simulated matches before producing a result he found fun .
  • He uses a project write-up as a self-imposed constraint: he won’t do a project without writing it up, which he says helps him finish work and leaves a useful record others can read .
  • For coding-model selection, he considers Terminal-Bench a reasonable signal but says spending time with a model has no replacement; he recommends settling on a good default rather than chasing a roughly 3% gain, while watching for new capabilities that unlock different work .
  • Agents reduce the need to memorize low-level trivia, not the value of understanding how the stack works; Simon says that deeper understanding helps developers use agents exceptionally well .
Simon Willison: AI Agents Made Me More Productive. So Why Am I Working Harder?
Logan Kilpatrick
Profile

For coding-agent evaluation, check whether the task outcome is achieved rather than requiring one implementation: coding can reach the same result through different implementations, and the panel describes coding verification as comparatively cheap. For simpler tasks amid increasing coding use, the panel says the expected bar is that the model “should always work,” since users will not inspect every line.

Evaluating the Frontier with Google DeepMind’s Gemini Team (Panel)
Jerry Tworek
Profile
  • Tworek recommends an automation-first posture: push agents slightly beyond their current limits and expect failures (“slop”) to reveal where automation breaks. He says current agents remain weak at high-level research and software/system design, and that he has trust issues with them.
  • Manage agents like a team: act as their “CEO,” while keeping responsibility for monitoring whether they do the right work and use resources appropriately.
  • His improvement thesis is experience-based learning rather than “whack-a-mole” fixes: look for agents becoming more successful at similar work over time.
CEO of your agents
Theo - t3․gg
  • In Theo’s hands-on evaluation, GPT-6.2 Astra was a strong code reviewer, catching issues other models missed without fabricating findings, and its computer-use actions felt fast because it anticipated and batched operations. It also proposed a reverse-proxy approach Theo estimated could save T3 Code up to $25,000/month if implemented. Despite these strengths, he planned to use Astra mainly for code review or possibly a model council rather than replace Opus 5.5 for his development work.
  • For long-running work, Theo reported Astra asking the human about obscure subagent details and blocking itself, then producing explanations so opaque that he switched to Opus 5.5 for readable summaries. In his TS-to-Rust project, Astra spent seven days without improvements, then four more with no changes; its size-focused work made the compiler 56% slower, while Opus later made it 50% faster in two days. He also reported Astra wasting resources and filling a mostly empty 8-TB drive with disposable files. Theo said later system-prompt changes improved one explanation, but it remained inadequate.
  • In an experimental game-level project, Theo’s provisional division of labor was Astra for general design and 3D work, with Fable for overall game design and cleanup; he considered using Astra as the playtester and Fable as the implementer.
Astra 6.2 Is Unusable
Latent Space

Anthropic’s disclosed usage mix suggests model recency matters: about 80% of usage came from models released in the prior six months. The discussion also notes that API-based workloads can shift readily to a rival model that pulls ahead, while first-party apps with scheduled tasks may be harder to switch away from.

⚡️Anthropic: Know Before You IPO — Tanay Jaipuria, Wing VC
Simon Willison

Simon Willison says he sometimes runs a “pelican” through a stealth model on OpenRouter to guess which lab built it . A reply describes fraudsters using that kind of model probe to gauge which model Claire Vo’s app uses; Willison says he had not realized the technique would be used that way .

[@webology](https://x.com/webology) [@clairevo](https://x.com/clairevo) Totally makes sense though, sometimes I'll see a stealth model on… [@clairevo](https://x.com/clairevo) [@simonw](https://x.com/simonw) Claire brings up the first case I have heard of your Pelican test bei… [@webology](https://x.com/webology) [@clairevo](https://x.com/clairevo) I talked to her about this in person at the Microsoft Ignite conf…
David Heinemeier Hansson (DHH)
Profile
  • DHH says he used the model transcribed as “Kimmy K25” as his primary driver for over a month, valuing its speed of about 200 tokens per second versus 30–40 tokens per second for frontier models at the time.
  • His Linux distribution project has agents integrated; he described using them to help make themes, apps, or plugins.
GOSIM 2026: the DHH keynote
swyx
  • In a one-week trial, swyx handed Devin Bots AEO, QA, cross-platform app porting, personal-admin automation, and video-editing work; because the bots are cloud-based, he also moved from managing local agent worktrees to a cloud workflow. He describes them as always-on but tuned for ROI, responsible for outcomes rather than individual tasks, able to manage and review subagents, reachable through text or Slack, and reporting through interactive Devinapps instead of long chats.
  • A reusable setup is to give a persistent, project-specific bot high-level context and responsibility for a goal: Devin Bots surface inter-session context by default and keep individual memory and identity, which can reduce repeated prompting on subsequent tasks. The product’s guidance is to delegate goals and responsibilities rather than isolated tasks.
devin bots are multi-agent teams done right: - always on, proactive (but tuned for ROI) - own outcomes not tasks (but can collab w others… Bots are super cool and if you haven't used them yet, you absolutely must try them. Create a bot to manage a goal or a project that you'l…
swyx

At AI Engineer Code Summit, Poteto (@poteto, identified as @SpaceXSI) recommends turning a repeatedly corrected mistake into a deterministic check and keeping the codebase “worth copying.” The talk is presented as explaining how the Grok bot team ships quickly without compromising quality.

If you keep correcting the same mistake, make it a check. [@poteto](https://x.com/poteto) of [@SpaceXSI](https://x.com/SpaceXSI) on turni… my AIE talk is up! come hear how the grok [@bot](https://x.com/bot) team ships so quickly without compromise on quality [https://youtu.be…
swyx

Netlify is repositioning for the AI era; it reportedly had a first day with 100,000 new signups, most new to AI and building software. Swyx raised open questions about why these builders choose Netlify over Vercel, ChatGPT sites, Anthropic artifacts, or Cloudflare, and how nontechnical users’ account growth and potential graduation from the platform develop.

my first startup and first love [@netlify](https://x.com/netlify) is refounding for the ai age with some very impressive growth stats des… Last week, Netlify enjoyed its first day with 100,000 new signups, most of them new to AI and building software. After taking 5 years to …
Riley Brown
  • Brown uses Claude Code’s document and design artifacts as a writing pipeline: he keeps writing in a pinned thread, asks Claude to draft in his voice and create matching visuals on a design board, then revises images with comments that send feedback back to Claude.
  • To connect an agent to business operations, he supplied Revolut Business API docs and asked Claude Code to build a custom MCP server for balances, transactions, cards, team, and payment drafts. The agent can analyze spending and prepare invoice-matched bulk payments in multiple currencies; Brown reviews the draft and sends the money himself.
  • His app-prototyping path is to turn an artifact mockup into an internal tool with a GitHub repo, Convex database, and Vercel deployment. He reports building a Final Draft-style script app in three prompts plus two changes, including collaboration, comments, image insertion, and PDF export.
  • He also uses Claude Code’s access to local files to generate a visual storage report and iterates on its charts through comments. He notes that bypass-permissions mode can give Claude Code broad access to the computer, making permission scope relevant to this workflow.
Claude’s NEW Artifacts Are a Superpower (How I Use Them)
Simon Willison's Weblog

AI agents translated 185,000 lines of C into 340,000 lines of Rust, while curl merged more than 300 upstream commits in a little over three weeks, including over 200 bug fixes and fixes for some vulnerabilities that the rewrites lacked; Daniel Stenberg says curl made roughly 200 more fixes the following month and questions whether the clones can keep up. This is a concrete caution that agent-generated rewrites may lag continued upstream maintenance.

Quoting Daniel Stenberg
Simon Willison

Simon tried Gemini 4 Neon with a prompt for a Three.js animation of a mech-version of Yonkers stomping around Manhattan, Queens, and Brooklyn using OpenStreetMap data; the generated result is linked in the post.

I also tried "Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and quee…
swyx

In Simon Willison’s GPT-6.2 Astra and Sol comparison on a creative-coding prompt—make a Three.js animation of a giant mech stomping around Manhattan, Queens, and Brooklyn using OpenStreetMap data—swyx said Sol won.

Here's GPT-6.2 Astra and GPT-6.2 Sol for "Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stom… [@simonw](https://x.com/simonw) dare i say sol won here
Kent C. Dodds 🐨

An image caption identifies Fedor Korotkov of Agent Infrastructure presenting a talk titled “We let agents hack our infrastructure” at an OpenAI event; no workflow or implementation details are given.

I'm at the [@WorkOS](https://x.com/WorkOS) office in NYC for [@meganreyno](https://x.com/meganreyno)'s [https://infra.nyc](https://infra.…
ThePrimeagen

ThePrimeagen is asking Omarchy users to help with a new round of bug reports through bug-or-not.trm.sh . The tool now offers skip and feedback options; he says it does not log or store user email or other information .

Calling on all Omarchy users who want to help A new round of bugs for omarchy, could you please help? [https://bug-or-not.trm.sh](https:/…
Kent C. Dodds 🐨

Luke Parker proposed deleting all repository tests and banning them in AGENTS.md, arguing that Opus 5.5+ can infer intended invariants from the code and that tests duplicate logic while adding agent compute and task time. Kent C. Dodds shared the proposal with the comment, “All my competitors should do this.”

i deleted all the tests in my repo and banned them in AGENTS.md tests were always just a second, worse copy of your logic. a leaky imitat… All my competitors should do this [https://x.com/LukeParkerDev/status/2122812034584641662](https://x.com/LukeParkerDev/status/21228120345…