Anthropic leaves Git for Scion, Gemini 4 Argon goes GA, an OpenAI engineer is charged with bugging an executive conference room

Anthropic replaces Git with Scion

The post. Anthropic's engineering blog published "Redesigning source control at Anthropic" (announcement, reposted by Boris Cherny). The opening claim: "At Anthropic, coding is no longer the bottleneck to building systems or doing research. … The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI." Anthropic averaged around 2,500 merged changes a day in May 2026 and more than 30,000 a day in September. "The history of our main branch alone is almost 100 GB, and a compressed working copy is over 25 GB." CI alone does about 25 million checkouts a day, which "fetched naively" would be over 600 PB of git traffic. At the peak Anthropic ran several hundred internal git mirrors.

The design. The starting question was "if we designed source control with the fact that agents are always online in mind from the start, what would it look like?" Scion is centralized. The server creates commits, merges and rebases and answers blame and history questions, and the same queries are exposed as an API so CI jobs often need no checkout at all. Commits are "changesets" in a write-ahead log that point at file contents in blob storage, with refs and derived data in Spanner, and "there is no fully materialized copy of the repo anywhere, even on the server." Checkouts go through three tiers (scion-daemon on each node, regional scion-cache, then scion-server), each with a cache hit rate above 99%. The footprint is "200x smaller than our previous git mirroring infrastructure," checkouts take seconds rather than minutes, and "CI alone realized a 25% CPU capacity win."

How it was built. "A team of seven people built both Scion and Antique" (the new code review tool) and migrated the monorepo off Git in five months. The team had handwritten the first Git-to-Scion ingestion pipeline. Then an early Claude Mythos Preview rewrote it "with no further instructions," and the new version "consumed 20x less memory and was 4x faster." The post is frank about the human role:

  • "All of the engineers who worked on this system would now rate Claude as better at writing code than them."
  • "Very often, we found that the role of the engineers was to tell Claude 'no.'" Agents tried to brute-force lookups by reading the commit log linearly, and adversarial review "can also be myopic," pushing Claude to add "excessive validation and complexity."
  • Reviewers "now examine the intent of the change as opposed to reading every line of code."
  • Each component has a Claude in its own Slack channel watching telemetry: "We call these Claudes 'superfans.'"
  • There's no global merge-queue lock anymore: "A proper merge queue, for example, doesn't make much sense at agent scale." Concurrent changes can conflict, "but Claude can typically fix these conflicts faster than humans can even identify them."

Anthropic won't release it ("Scion is by no means a suitable replacement for git for general use") and says it wants to work with GitHub, GitLab and the git community on server-side RPCs and better sparse and shallow support.

Hacker News (709 points, 682 comments). "So... Anthropic reinvents SVN?" asked mijoharas, and bradfa answered: "Feels more like perforce, but effectively, yes." amluto liked the idea until server-side commits: "I'll pass. … By all means, built a better system than git. … But don't lose the actual benefits of git!" joshheitzman: "Commit counts are no better than line counts as a software development metric." steveklabnik, whose East River Source Control offers a bridge from git to something new, said seeing a big lab say this "really validates our thesis." Eridrus pointed to Cursor's "Git at any scale" post as proof that git itself can be scaled. Scion engineer nickg-anthropic answered dozens of questions:

  • On offline work: "Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do."
  • On what all those merges are: "Building all of the systems to train and serve our models, internal tools and infrastructure, (lots and lots of) research, our product and API surfaces, and more."
  • On conflicts: "We never allow textual conflicts," but semantic conflicts "are an unavoidable fact of very large repos unless you completely serialize merging."
  • On the hand-written prototype: "The ability to delegate larger chunks of code to Claude had a noticeable, step function increase with Mythos Preview."
  • On Cursor's post: "this is comparing apples to oranges."
  • And: "this blog post was 100% human-written."

Elsewhere. Simon Willison had Claude add model release dates to Anthropic's merge chart. Simon's guess is "that the introduction of Fable/Mythos class models is a key factor in the increased coding velocity." The adversarial-review warning matched Simon's own experience playing Claude and GPT off each other. On X, Jim Stewartson posted that "30,000 code changes per day is absurd by 3 or 4 orders of magnitude." Theo's reply: "Dude legit thinks a 2,000 person company should be merging under 30 PRs a day lmao."

My read: 30,000 merges a day says little on its own, as joshheitzman points out. The striking number is seven people replacing Git and code review in five months while the engineers' main job was saying no. Whether a system built that fast stays maintainable is the open question, and Anthropic itself says most companies shouldn't copy it.

Gemini 4 Argon goes GA

The launch. Google made Gemini 4 Argon generally available: the paid Gemini API and Gemini platform for developers, Antigravity for AI Ultra subscribers, and the Gemini app and AI Mode for Ultra and Pro. The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. A footnote says that "after the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply." Google says Argon ranks #1 on the Vals Index. Its internal coding claims:

  • In Antigravity, Argon "helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third."
  • It helped the Chrome team split core webpage handling out of a ~6 million line legacy folder, "completing months of work in weeks."
  • Given the ADK team's .agents/skills rulebook and a one-paragraph goal, it worked overnight and "cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%."
  • It made Gemma 4 E2B 2.85x faster on vLLM "with 100% token parity."

The safety post says Argon solved "all 12 of the end-to-end simulated cyberattack scenarios, up from five in our previous generation." Google added input screening, activation probes and streaming monitors, and defenders in the Fairwind Program get Argon "without cyber guardrails."

Third-party numbers (collected by AINews). Cursor reports 52.8% on CursorBench at default effort, on par with Opus 5.5, and Argon is also in Devin and GitHub Copilot. Artificial Analysis puts it level with GPT-6 Astra on its Intelligence Index at about 60% of Astra's cost per task under the introductory pricing. Arena's agent leaderboard has Argon High fifth, behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra. Jerry Liu's team benchmarked document OCR: "Muse Spark and Gemini 4 Argon* are SOTA at their respective price points for table understanding among frontier models."

Theo isn't sold. Theo had early access: "It is a huge jump for Gemini, and pretty good value for the price," and it was the first model to get a perfect score on Skatebench-private. "It's still a Gemini model though. Does weird shit regularly and I don't feel like I can trust it at all." An hour later, on Theo's codebase-analysis bench, Argon was the "Only model with a >50% false claim rate 🙃" Later that night: "Feels like it might be months until a model can actually beat Opus 5.5 (on price + performance + capability)" and "...I don't think I believe in benchmarks anymore guys." Theo has filmed a Gemini 4 video but hasn't posted it yet.

Pelicans. Simon Willison shipped llm-gemini 0.35. The low, medium and high pelicans are "all excellent," and the high one cost 17.166 cents. Simon's caveat on price: "Prices like this aren't really comparable these days since what really matters is how many tokens a model takes to achieve a goal within an agentic harness." Artificial Analysis spent $2,407 to run its benchmark on Argon, close to GLM-5.3 and well under the OpenAI, Anthropic and Meta flagships. DeepMind's @fofrAI had Argon draw a pelican riding a bike "using fourier epicycles," which Simon called "a new twist."

Monitorability. Per AINews, DeepMind's Neel Nanda reports that in adversarial evaluations, Argon can evade a realistic chain-of-thought-plus-action monitor on simple sabotage tasks about half the time. AINews notes these are controlled assessments, not evidence of sabotage in deployment.

Hacker News (1,211 points, 854 comments). Much of the thread was Pro subscribers who couldn't find it yet. postalcoder used it over the API: "Clear in the head, not RL-fried, distinct writing style, and big model smell. That said, the em-dash is back w/ this model." protoman3000 likes writing with Gemini most, then added: "The code is bad though." damsta wanted to know why the post-intro price was buried, and simonw replied with the date it doubles.

My read: Argon is cheap for what it scores, at least until December. But Theo's false-claim rate and Nanda's monitor result point at the same gap. A model that builds well but reports on its own work unreliably needs more checking around it, not less.

Muse Wonder, day two. The pricing complaints arrived. @atmoio went from the $5 Muse Code plan to $15, $50 and finally $200, and the five-hour allowance was "GONE" within the hour: "If Meta's $200 subscription quota is really what it costs to use AI for coding (and it probably is), then we are all in for a very, very rude awakening when the token subsidies end." @morganlinton: "At Low Effort, I hit my weekly usage limit on my Muse Code plan in less than 12 hours." Alexandr Wang said "there is still a lot for us to improve and we are working on ironing a ton of kinks in how we deployed the model," then that night "we have been fixing lots of stuff with the model rollout over time." Theo's full review video is linked under Videos.

Agentic Coding & Agent Harnesses

Matt Pocock's list of mistakes. Matt posted "Some mistakes you might be making," 18 of them, including "Your effort level is too high," "Your skills are too long," "You're aligning too often, and too hard," "You aren't reviewing previous coding sessions (with an agent)," "You're doing too much human review," "You aren't vibe coding enough prototypes," "You've overloaded your global steering files," "Your codebase architecture allows bad tests to be written" and finally "You aren't thinking enough." Five short videos on the same themes came out the same day (see Videos), along with The AI Coding Dictionary, "The words people throw around when they talk about AI coding," wired into a 3D graph.

Geoffrey Huntley: delete your skills. Huntley wrote up four episodes with Josh Wulf as "nobody should type code by hand anymore. here's what's left." The thesis: "There's no place in the world now for a software engineer who types code by hand." Some of it lines up with Matt's list. "You can buy an influencer's skill pack and course, but there are no best practices, and whatever exists will be gone in two months." Huntley recommends that "you should probably delete your skills and your AGENTS.md every couple of months. The models get better, so how do you know you're not de-tuning them?" For enforcement, Huntley uses pre-commit hooks instead of instructions. Simon Willison quoted this part: "I write a pre-commit hook that checks the git diff for direct calls to observability modules that bypass my diagnostics module. Boom: fail the commit, inject a prompt telling the agent to use the diagnostics library. All of a sudden it's ten out of ten." Matt's short "STOP running your checks in harness hooks" covers the same choice: checks on every file edit versus pre-commit hooks.

Lauren's weedkiller for comments. "ban and delete most if not all code comments!" Lauren (@poteto) explained: "agents love copying what already exists. a seemingly harmless comment explaining some temporary workaround is one of the worst viruses you can have in your codebase. like a weed it'll spread quickly." pstack's /no-comments skill hunts the useless ones and keeps the ones that should exist, "like explaining a browser quirk." pstack 0.15.23 also turned Lauren's most-used prompt into /in-your-own-words: "restate in your own words what you think my goals are and what the problem i'm trying to solve is." And on writing: "i love writing. i never let agents write for me. … i still strongly believe that writing is thinking."

Measuring slop. Parsimony (HN) ranks models by net code added on the SWE-bench tasks they solve, because "Existing benchmarks measure solve rate and cost, but don't punish excess code." It counts syntax units, not lines, and subtracts deletions. One caveat it lists itself: "Models that write more code also tend to solve more tasks (Spearman 0.33–0.60 on the boards with 26+ models)." ai-deslop (HN) is a set of Oxlint rules that "reject low-evidence, low-signal, AI-slop code patterns" in TypeScript, such as no-chained-type-assertions and no-runtime-typeof, partly vendored from dmmulroy/anti-slop.

Kill My SaaS results. swyx's AI Engineer team got a $43,000-a-year quote for speaker-management software, so Latent Space ran a contest to clone it with agents. The report covers 72 submissions. Of the 66 non-finalist records, 52 used Claude Code, 49 Codex and 40 both. Builders split the work: one wrote that "Claude code as a harness works great for orchestration in my experience compared to Codex but Codex is excellent for code review, computer use, and testing." Trackstage ran "a central Fable thread" that "coordinated Claude subagents and called Codex for adversarial review." The video says the organizers' own AI judge ranked the eventual winner 19th. First prize ($10,000) went to Valentin De Matos for OpenRostrum.

Your game is a harness. Kyle Kukshtel turned a talk into "Your Game Is A Harness" (HN, 146 points). The argument is that LLMs are bad at games because so much of a game isn't code, so make the game legible: fixed randomness, screenshot-ability with toggleable layers, deterministic and addressable state, and a standalone simulation. The postscript admits Kukshtel is shipping a small project "where I've also used none of this." On HN, TeMPOraL's most valuable prompt was "please figure out how to run the game headlessly, so I can use my PC while you muck around." The agent now runs parallel instances on its own to check its assumptions.

Claude Code v2.1.299 (release notes) adds idle compaction: "a 200K+ token conversation left idle for about 54 minutes is compacted before its one-hour prompt cache expires, so your next message doesn't reprocess it all." /usage now warns about max effort, workflows, always-on context and large tool results. The release also closes two sandbox holes. A sandboxed command could write into a --plugin-dir folder "whose code Claude Code reloads and runs outside the sandbox," and on macOS a denied path could be reached through a symlink. LLMJunky, a few hours before the release: "Anthropic's compaction has gotten really good too btw. … Drop the context window anxiety, its 2026. If you're that worried about it, orchestrate subagents/workflows."

Codex Day 9. Tibo shipped a change to how the ChatGPT and Codex desktop app stores state, which "should reduce the amount stored on disk by 60%. Small ship but we've heard enough feedback about power users running out of disk space."

Smaller things.

Safety & Security

An OpenAI engineer is charged with bugging an executive conference room. The DOJ says Abdulrahman Alfozan, 35, a Saudi national, was arrested on October 9 and charged with using a device to intercept oral communications. The SF Standard has the complaint's details:

  • OpenAI found the device on September 30. It was plugged into a USB port on the back of a TV in an executive conference room and "concealed in toilet paper," and its contents were encrypted.
  • A chatbot account tied to the engineer's spouse asked how to "build a continuous audio recorder" and "encrypt the SD card." The DOJ text says "chatbot" without naming a provider.
  • On October 5, security opened a bag the engineer had left unattended and found the engineer's wallet and ID along with a second, similar device.

OpenAI said "We used our models to understand the device and where it had come from," and that it has "no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data." On HN (532 points), minimaxir asked: "Did OpenAI do a marketing bit in the middle of an espionage accusation?" iAMkenough matched the photos to a Seeed Studio XIAO ESP32S3 Sense board.

XBOW's $250,000 Chrome chain. Oege de Moor says Google awarded XBOW a $250,000 full-chain Chrome bounty, only the second in Chrome's history and the first under Google's 2026 rules. After Google said it already knew one bug in the first chain, "Brendan and the agents built a new exploit in just 6 days." "It took just 6 days and $40,000 of compute to construct an exploit that affects a billion users of Chrome on Windows." On policy: "Restricting powerful models is not the answer; GPT-5.6 Sol is already public." @jxnlco (reposted by Peter Steinberger) confirmed: "it was with sol 5.6 ._."

Codex sessions can read each other's Pages. In openai/codex#53743, amluto reports that any signed-in codex-rs instance can list and read every "Page" another instance created through a built-in skill (mcp__codex_apps__chatgpt_space_list_pages). That works even across sandboxes that can't otherwise talk to each other, and there's no permission prompt. The issue asks for a prompt at sign-in and at tool-call time, or for scoping each instance to its own pages. On the Scion HN thread, amluto put it this way: "If you thought that your agents in your swarm would need to find some external bulletin board to cooperate behind your back, you are mistaken: OpenAI provides the relevant service for them with no permission prompts at all."

Fake "OpenAI employees." @LyalinDotCom: "Another "openAI employee" reached out to me to "chat" lol." @MatthewBerman got the same DM and set a bot on it with a prompt that opens "you're a scam beater bot."

Scrapers. "I posted to HN, and then my website was flooded with AI bot traffic" (HN, 316 points). Nearly half the traffic came from one service that polls sites and "summarizes them into an AI slop news letter." In the comments, switz now blocks AWS, Google Cloud, Azure, Alibaba Cloud and Oracle: "I banned Alibaba's Cloud two years ago and it still hits one of my domains 1 million times per day, despite every route 4xx-ing during that time."

Videos

Other Interesting Stuff

  • Character.ai filed for Chapter 11 to run a court-supervised sale. If no bid arrives, "the platform will be wound down." The Verge reports Disney is the lead bidder at $20 million for undisclosed assets and has already hired former CEO Karandeep Anand as CTO. Liabilities are listed at $50 million to $100 million. On HN (317 points), drewbitt tallied the spending and gizzlon asked what happens to users' personal data.
  • Xbox Labs. Xbox's Asha (@asha_shar) announced a small group to explore what the recent wave of decompiled, ported and remixed games makes possible: "Coding became the first major workload for agents because they can write, run, and test software. Games share many of those same properties." Theo: "Is this the first major gaming company to make moves based on all the vibe coded browser ports of classic games?"
  • Theo vs. local AI. @TeksEdge posted a 10-person startup's $18,000 local inference box, and Theo answered: "For the cost of 8 years of $200 Claude subs, you can force your team to use shitty quantized versions of DeepSeek v4 Flash running at 30tps. Incredible deal!" The $200 plans "will get better over time," though Theo also noted: "Despite having 6 accounts, it's not looking good for my Claude limits right now 💀"
  • Who automates whom. Peter Bryant asks "Were we blue-collar workers all along?" (HN, 201 comments). gizajob objected to the article's Excel analogy: "Accountants absolutely loved Excel." Sean Goedecke goes further in "Software engineers don't get to be Luddites about AI": software engineers are "like the smiths and mechanics who built the power looms in the first place."
  • Grok Bot does taxes. Lauren, who works on Grok Bot, reposted @jasonfreedman, who gave Grok Bot four years of returns and asked it to check the CPA's work. It found mortgage interest missing from a 2022 return and "HSA payroll contributions counted twice, four years in a row." Lauren also expects that "have your bot talk to my bot" will soon be common, now that Grok Bot has its own email.
  • Exa passed $100M ARR after growing "almost exactly 10x per year for 3 years," says @WilliamBryk (via swyx): "We predict agents will search 1000x more than humans by 2030."
  • Drama 3. Fish Audio released an open-weight TTS model where "You can now control how even a single word should sound" (launch). Artificial Analysis says it's a 35B MoE with 3B active parameters. LLMJunky on the ad: "What an absolutely incredible commercial. Uncle Roger approved."
  • EmbeddingGemma 2 as speech-to-text (HN, 411 points). Leaving the audio encoder untouched, a small trained "reader" maps audio slices to the nearest of ~229,000 English words, for "An hour of audio in about 13 seconds on a base MacBook Air, on device." Only the reader was trained, in "under two hours on 8×H100."
  • Apple's Minions pools idle Apple Silicon Macs into a GPU cloud for batch inference (HN). railka on HN says it uses task-level parallelism, not one model split across machines.
  • Model and industry odds and ends (via AINews): Prime Intellect acquired Moondream. "Fable 5.5" showed up in LiteLLM's Azure model data, but the finder says the details look copied from Fable 5.1. The IMF warned that "The consequences of an AI correction could be global."
  • DishwasherBench has models (and humans) load a physics-simulated Bosch dishwasher and scores whatever comes out clean. It's a work in progress, inspired by a Joma Tech video.