Gemini 4 Carbon takes #1, Claude Code's hidden summary bill, GitHub goes multi-model, Opus 5.5 makes fun games, and Thariq on sad engineers

Agentic Coding & Agent Harnesses

Claude Code's status line was eating your tokens

The top Claude Code story on HN (315 points) is a gist from Flamerix AI: "Every ~30 seconds, for every background agent, Claude Code sends that agent's entire transcript to the agent's own model (usually Opus) and asks it to 'describe your most recent action in 3-5 words'." The answer only feeds the agents panel. Across one week of their OpenTelemetry data, these agent_summary calls came to 16.4% of total spend, more than the main conversation (15.7%), and 19.7% of all API requests. There is no setting to turn them off, so the gist ships a Bun preload that intercepts the timer without patching the binary.

edwinarbus, who works on Claude Code, replied that the summaries only run for background agents that are actively working. Across users who send telemetry, "agent_summary was about 2% of cost at API list prices and just under 4% of requests", and about 6% in the terminal. "That said, we agree that's a lot for a status line." A fix for summaries that miss the prompt cache is out in 2.1.311+, and a cheaper mode is in the works that "sends short notes to a smaller model instead of the full transcript". Another commenter measured "~8% at most" and found that workflows (Ultracode) weren't affected. Another replied: "8% still seems pretty indefensible, especially if you can't turn it off."

Separately, am.will noticed that Ultracode is now a toggle at any effort level in Claude Desktop ("like crack… I'm two days clean"). One reply said it launches "20–30 agents—which drains the 5-hour limit for the Max 20x plan before the task is even finished".

GitHub Universe: Copilot Fusion, Projects, and agents tripling commits

GitHub's Universe keynote recast the platform as an "agentic software delivery system":

  • Copilot Fusion (GA) replaces picking a model. It routes each prompt and can orchestrate several models (Single, Cascade, or Critique, where one model reviews another's draft), at Standard/High/Max effort. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, "within two points of Claude Code with Opus 5.5 Max", at about half the cost and tokens and about 40% less time per task. Admins control which models it may use.
  • Copilot Projects (public preview): give the Copilot app a goal that spans several repos plus Slack and Figma context, and "Copilot runs as many agents as the work needs" and brings back PRs. There are also project automations and computer use on macOS and Windows, with approval before Copilot controls an app.
  • Octoverse. Simon Willison, there on a GitHub Stars ticket, summarized the report: merged PRs "tripled in less than a year to 220 million per month", commits more than tripled to 7.38 billion a month (mostly in the second half of 2026), and new repos doubled to 30.3 million a month. dax posted the same curve for the OpenCode team: "here are our teams commits since january… something is happening". Theo replied "Please tell me you guys stopped using Argon before that."
  • Matt Pocock in the keynote. Jason Lengstorf spotted "@mattpocockuk getting singled out in the keynote". Matt explained the joke: a Satya Nadella post describes building an app with "a single prompt + skill (/drill-me)", which Matt reads as Matt's own skill under "a rather bawdy name".

Simon Willison: October is the inflection point for vibe-coded games

Simon argues that as November 2025 was for coding agents, "we've reached a new inflection point in October 2026, and this time it's for vibe coding computer games" (tweet, 1.2K likes). Opus 5.5 improved at 3D modeling, music and, most of all, game design: "Then I tried telling Claude Opus 5.5 to 'Make it a LOT more fun'... and it did." Simon made four games without reading the code. Tree Appreciator came from one prompt in plain Claude chat. Emberwick is a 15-minute three.js RPG whose ending left Simon "emotionally affected". None took more than four hours of Simon's own time. Simon's impression is that GPT-6 Astra "isn't quite there yet". A section titled "Four times as insulting" acknowledges that these games combine AI art, music, writing and code. The advice: "Build a mobile-friendly three.js game where..." and switch to Codex or Claude Code once it's past proof of concept.

From the replies: @pwnies's group built the game at the start of a LAN party instead of choosing one. @traits_reality pushed back that "it's pretty awful at game balance… it will decide on a metric and overfit on it so the result is plausible but unfun". @ArmataGame described a check where "a fresh agent that only sees screenshots signs off every visual change".

Theo: T3 Code goes multi-machine (Days 24–29)

Theo shipped six "days" in about three hours:

Video: "How To Give Claude Unlimited Usage": "today I'm going to show you how to run Claude for WEEKS". Theo calls it "a chaotic spiral into hill climbing", the start of a series, after "claudemaxxing a bit too hard lately".

Matt Pocock: the black hole of success

Matt proposed (700 likes) that the old "pit of success" was limited by a social contract: "You couldn't make the pit too steep, otherwise the humans would get annoyed… Agents don't feel annoyed, so you can crank up the pressure to infinity… Make great code inevitable." A worry from the replies: "if the agent can't cross back out, it also can't tell you it's stuck." Another predicted that "max out the checks for agents and i'd expect your engineers to start skipping the hooks."

Matt's example: "Debating just straight-up banning useState in my React apps" (1K likes, 207K views). The replacement is useReducer, which fits Matt's "Banger PR" written with the /pr skill: "a great example on how to contain complex frontend state in a reducer".

Matt also finished filming "Software Factories for Real Engineers", starting November 9 for two weeks. The pitch: "Loops, YOLO mode, dark factory. That's a recipe for a slop cannon", so "slow down to speed up" with automated review, codebase design, DDD, mutation testing and feedback loops. You leave "with a software factory spec, ready to hand to an agent." Matt says it's already on track to be the biggest course of Matt's career.

dax on open-model demand and two years of coding agents

After yesterday's AA v5 results, dax (OpenCode) added context (1.5K likes). The top inference providers chasing enterprise open-model business "are all roughly the same size as us. and at best there's 5 of them". Few companies can run reliable inference themselves, and "you just have to look at products… how many products exist with a lot of traction that can use open models?" Pushback came from a Russian bank/marketplace developer whose employer self-hosts GLM and DeepSeek, and from @jamiequint: "Your incorrect assumption is that most inference is coding harnesses. It's not."

Earlier, dax wrote: "it's getting close to 2 years of coding agents being a thing… the majority of discussion on here on where they're going, what's cutting edge, what matters has been primarily wrong… you had to actively ignore what was going on here to win in this space." The team also shipped hokum, Wayland notifications built on their new native UI framework, and a proper release of hex, local-first voice dictation. dax also warned that with supply chain attacks rising, "AUR really has to become more sophisticated or get replaced."

Quick hits

  • Peter Steinberger: OpenClaw's web UI now works offline with websocket compression and reconnects, tested on a flight to Vienna. On building your own agent: "Nobody can tell you no or limit what model you use", though "I use dots a lot at work." Peter also joked that you can tell human bugs from agent bugs: "Someone at Disney still writes code by HAND?" One tell Peter pointed to: "all caps title, agents would never".
  • Cursor's Agent SDK (noted here on 10-27) got its HN discussion (122 points). It's the framework behind Bugbot and Security Reviewer: define agents in code and deploy them with one command. One Cursor team built a PR checker against their lessons-learned knowledge base "in about a day". HN mostly complained about the name ("Open AI and Anthropic already have 'Agent SDK' brands").
  • Harness overhead: per AINews, the default system prompts of 30 open coding harnesses range from 235 tokens (goose) to 11,974 (Gemini CLI), before tool definitions.

Claude & Anthropic

Five exploit chains from an unreleased model

Anthropic published details of five attack chains found by "an unreleased, internal research model". In each case Claude wrote the exploit "after initial prompting":

  • Chrome: one-click RCE outside the sandbox (V8 heap corruption, a V8 sandbox escape, a CacheStorage race, and a use-after-free in FedCM). "62 CVEs have been assigned and patched as of Chrome 156", all critical or high, including 13 sandbox escapes.
  • Safari: a one-click same-origin-policy bypass that could read history, cookies and logged-in pages.
  • Android: extracting photos from a stolen, locked Pixel 9 by chaining three bugs, partly found by deobfuscating Google Play services.
  • Windows: a remote crash of any machine serving native HTTP/3 (an out-of-bounds read in the kernel's QUIC stack).
  • Linux: zero-click remote code execution in the kernel networking stack, now patched upstream.

Two lessons: "Test-time compute scaling and scanner diversity substantially increase how many vulnerabilities you find", and scan at PR time. The security-guidance Claude Code plugin is public. The expanded Cyber Verification Program gives vetted professionals Opus 5.5, Sonnet 5.5 and Mythos 5.1 with reduced blocking.

Also from Anthropic

  • Utah pilot. Boris Cherny shared (2.5K likes) that "Utah gave Claude Code to 34 developers across 10 state agencies. 80% had a working environment within 30 minutes, 71% reported saving 20+ hours by the end of the four-week pilot." Top reply: "the other 20% are still waiting for IT to approve the npm install".
  • Smart reports. Mashable reports that Claude Enterprise admins can now generate reports that read a team's transcripts across chat, Claude Code and Cowork, group them into workstreams, and show spend. Mashable's headline: they "tell bosses when workers are 'off-topic'". Ten reports a month are free.
  • Claude for Legal artifacts: a gallery of ready-made artifacts for contract review, regulatory horizon scans and cite checks.

Models & Benchmarks

Gemini 4 Carbon: #1 on AA v5, but not for you yet

Google announced Gemini 4 Carbon two weeks after Argon (HN, 906 points). It's pitched on cybersecurity and long-horizon work: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, and internally it has run "for days to persist through dead ends". Access goes first to "a subset of cyber defenders" through the Fairwind Program. The introductory price is $2/$10 per million tokens; after that it's $4/$20.

Artificial Analysis (4.6K likes) puts it at 59 on Intelligence Index v5, ahead of Opus 5.5 (58). It leads 4 of 9 evals, but on the Coding Agent Index it scores 59 with Antigravity CLI against Claude Code with Opus 5.5 at 62. It costs $7.76 per task against Opus's $10.04 and uses 29% fewer output tokens. Its hallucination rate rose from Argon's 4% to 11%, still second lowest (Opus 5.5 and Astra are at 50%). Google claims successful cheating on Terminal-Bench 4.0 fell to 0.0%, which an HN commenter called "probably the single most important thing to make frontier models usable right now". Per AINews, Carbon also tops Text Arena and Vending-Bench 2, but is fourth on FrontierSWE behind GPT-6 Astra, Opus 5.5 and Sonnet 5.5.

The objections are about access. "the chart literally says no publicly available api. so everyone's crowning a model they can't call yet". "would love a new rule: no api access, no spot on the index". On HN: "I have a girlfriend, you just won't know her because she goes to another school"-type energy. swyx opened Antigravity "for the first time in a long time" for an early review: "if google actually pulls this off it is officially comeback season with Meta, Grok and now GDM being cool again."

The open-weights argument from yesterday's v5 launch kept going. Theo: "Terrible day to be an 'open weight models are just as good as Opus 5.5' guy". am.will wrote "I want there to be Opus 5.5 at home so badly, but the narrative spinning is so unhelpful."

Also in models and evals

  • DAYJOB: Tax (Surge AI, HN): 50 professional tax assignments across 1,789 files. "Claude Sonnet 5.5 leads with a 10.8% pass rate", followed by Fable 5.1 (9.6%) and Opus 5.5 (9.2%). In one example, "Across five attempts, Astra never identified the deadline. Opus identified it in all five attempts."
  • Document Parsing Protocol: LlamaIndex open-sourced the 140-rule spec behind LlamaParse for things like split tables and checkboxes. Jerry Liu frames it as an evaluation rubric, since parse benchmarks "inherently reward a certain style even if other styles might be semantically correct". It comes with a Claude skill.
  • Qwenputer (HN, 392 points): a RISC-V CPU compiled into Qwen 3.5 4B's weights, using its Gated DeltaNet layers as exact RAM. It runs C and MicroPython inside the forward pass and handles the model's arithmetic. The same author built an embedded lie detector.
  • Nemotron 3.5 Transcribe: NVIDIA's open 30B MoE (~4B active) speech model for 40 languages, tied for #1 among open models on AA-WER.
  • Rime's prism (600K views) claims to be the first speech-to-speech model "you can't tell from a human in conversation", with 300ms responses. Top replies said otherwise: "Within 0.5 seconds of the AI talking, i could tell it was AI".

Codex & OpenAI

Personal Agents

Other Interesting Stuff