AI News Briefing

The Big Picture

Anthropic quietly published the most interesting engineering post of the year and almost nobody noticed, because it has no model in it: they threw out git. Not wrapped it, not sharded it — replaced it with an in-house system called Scion, after daily merges went from 2,500 in May to over 30,000 in September. Git assumes a client that works offline and holds the whole repo; Anthropic’s clients are agents and CI jobs that are always online and never need history. Checkouts now take seconds instead of minutes, with 200x less infrastructure. That is what agentic coding at scale actually costs — not better prompts, a different version control system. Meanwhile the walls went up everywhere else. Jamie Dimon said no outside agent touches a Chase account, full stop, no timeline for one that could. Eighty-plus publishers under SPUR now have a telemetry standard to bill for scraping. Amazon already shut Meta’s Muse out. The pitch for agents was that they’d go do things for you across the internet; the internet is answering, and the answer is credentials-or-nothing. Google also made Argon broadly available at half price until December, which is how a lab says the cyber-defender exclusivity was always marketing.

Claude Code v2.1.299 adds idle compaction to beat the one-hour prompt cache

A 200K+ token conversation left idle for about 54 minutes now gets compacted before its one-hour prompt cache expires, so the next message doesn’t reprocess the whole thing. Tunable via CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, or off with "idleCompaction": false. Also new: CLAUDE_CODE_DISABLE_FILE_MENTIONS so @ mentions pass paths instead of file contents, /effort and /fast in agent view, and /usage warnings for max effort and always-on context.

Two of the 131 bug fixes close real security holes: a sandboxed shell command could write into a --plugin-dir folder whose code Claude Code then reloads and runs outside the sandbox, and the server-managed settings request could carry a third-party gateway’s API key.

Anthropic replaced git with an in-house VCS after merges hit 30,000 a day

Anthropic migrated its monorepo off git to a system called Scion in five months, alongside a code review tool (Antique), code search, and source-control access control. The trigger: roughly 2,500 merges a day in May became over 30,000 by September, and the several hundred internal git mirrors propping that up stopped scaling. Main-branch history is near 100 GB; a compressed working copy exceeds 25 GB. Scion’s footprint is about 200x smaller than the mirroring infrastructure it replaced, CI and research checkouts run two orders of magnitude faster, and CI alone got back 25% of its CPU capacity. Not open source.

Anthropic’s argument is that git’s core assumption broke: it was designed for offline clients holding full repo copies, and Anthropic’s clients are now mostly always-online agents and CI jobs that never need local history.

Google drops the Fairwind gate on Gemini 4 Argon, halves API price until December

Thirteen days after unveiling Argon as a cyber-defenders-only release through the Fairwind Program, Google opened it to everyone via the paid Gemini API, AI Studio, Antigravity and Gemini Enterprise. Introductory pricing is $2 per million input tokens and $10 output, with 95% off cached input — expiring November 30, after which it doubles to $4/$20. Consumer access stays paywalled behind AI Pro and Ultra; individual Antigravity access needs Ultra. The API is now described as a combined 1M-token input/output limit rather than 1M output tokens.

Argon sits first on the Vals Index at 68.90%, ahead of Claude Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%), while scoring 53 on Artificial Analysis’ Intelligence Index — tied with GPT-6 Astra and Fable 5.1 but behind Opus 5.5 at 58.

Dimon bars outside AI agents from Chase accounts, with no timeline to reopen

On JPMorgan’s Q3 earnings call, Jamie Dimon said the bank won’t let outside agents operate inside its systems: “We’re not going to allow just any outside agent to come in inside the company because that’s just too risky for clients and for JPMorgan.” He named Meta’s Muse and Instinct and warned about “rogue agents.” A spokeswoman confirmed third parties cannot obtain customer banking credentials to log in on their behalf. Read-only balance access through aggregators like Plaid still works; sending money and paying bills does not.

The bank says there is no timeline for safeguards that would let agents act on accounts directly — while it builds its own agents internally and runs a Smart Cash pilot, with customer-facing AI planned later.

OpenAI engineering lead charged over recorder hidden in executive conference room

OpenAI disclosed that Abdulrahman Alfozan, an engineering lead, was arrested October 9 by the FBI and charged with intercepting oral communications. Per the DOJ complaint, a recorder wrapped in toilet paper was plugged into the back of a TV in an executive conference room in San Francisco. OpenAI says it dismissed him and contacted the FBI itself. It is a complaint, not a conviction, and the investigation is open.

OpenAI’s own statement is the tell on scope: it says it has no indications the individual inappropriately accessed or exfiltrated model weights, source code, or customer data — the three things a frontier lab actually fears losing.

NYT, AFP and The Economist join SPUR as publisher coalition passes 80 members

More than 40 organizations joined SPUR (Standards for Publisher Usage Rights) at once, including the New York Times, AFP, The Economist, The Atlantic and Brazil’s National Association of Newspapers — putting the coalition past 80 members in more than 20 countries. Founded in February by the BBC, FT, Guardian, Telegraph and Sky News, SPUR doesn’t license content itself; each publisher still negotiates its own deals. It is funded entirely by its members.

The leverage is the Content Telemetry Standard, finalized in early October after a June draft: it logs when a scraper fetches an article, how the text is cited or displayed in an AI answer, and whether a reader acts on it — turning “pay us” into a measurable invoice.