Gemini 4 Argon goes global 🌍, Hidden mic in OpenAI's exec room 🎙️, Anthropic moved off git 🌳

Subscribers get it in the Gemini app and AI Mode in Search, and Ultra subscribers in Antigravity, Google's agentic coding tool. Google says Argon tops the Vals Index of finance, coding, legal and tax work and lifted its own teams' agent-written code submissions by over 35%, with nearly a third fewer rollbacks. More Gemini 4 models are coming, and scientists can apply for one of 10,000 free six-month places on Antigravity's top usage tier.

OpenAI engineer Abdulrahman Alfozan allegedly hid the device, concealed in toilet paper and plugged into a TV, in an executive conference room he had no authority to access. The complaint says a chatbot account tied to his spouse asked how to build a continuous audio recorder and encrypt its SD card, and a second device turned up in his bag. The FBI calls protecting American AI innovation a top counterintelligence priority; he faces up to five years if convicted.

Subscriptions peaked at $2.8 million a month and total revenue at $5.3 million, and as of August it was still losing about $8-9 million a month. CEO Colin Adams says it cut costs by moving to cheaper GPUs, adapting open-weight models instead of building its own, cutting staff and adding ads. 404 Media argues the runway is now over for second-tier AI companies without repeated cash injections or an ad business to subsidise them.

CEO Jamie Dimon told analysts on Tuesday that letting outside agents in is "just too risky for clients and for JPMorgan", and there is no timeline for safeguards to allow it. Agents can still read data such as balances through connectors like Plaid, but cannot send money or pay bills. Dimon says customers who use outside agents should understand "there's a liability transfer", while JPMorgan pilots Smart Cash, its own tool for moving funds to maximise yield.

Agents connect through its MCP server, which offers 275 read tools and 54 that write or delete data, and sign in as the user, with that person's permissions. Workday says data-lake copies bypass those controls; in a demo, an agent reading one in Snowflake produced the CTO's salary. Product and technology president Gerrit Kazmaier says building directly on those tools works for only "a small set of use cases", pointing everything else to Sana, Workday's agent orchestrator.

Teams upload demonstration data, launch fine-tuning on compute the company manages, and run the tuned model on their robots through its cloud. Physical Intelligence says early-access partner Almond's Axol robot smooths rough edges off spacers for its own assembly using a model tuned over four days on the API, though some cloud-run deployments hit network problems. Access is by application: the company asks would-be users to tell it what they want to build.

Liquid says it generates 184 tokens a second on an Apple M5 Max and leads or matches larger rivals such as Qwen3.6-35B-A3B on five of eight agent and tool-use benchmarks. Its table shows Qwen ahead on coding, 66.2 to 51 on SWE-Bench Verified, and Liquid suggests adding retrieval or web search for knowledge-heavy work. The base and post-trained weights are on Hugging Face, with day-one llama.cpp and MLX support for running it on your own hardware.

A seven-person team built Scion in five months, an in-house system where agents and CI jobs fetch files on demand from always-online servers rather than holding full copies. Anthropic reports it needs 200 times less CPU and memory than its old git mirrors, with checkouts in seconds rather than minutes. It calls Scion no general git replacement, but expects teams using AI coding tools to hit these problems more often within six to 12 months.

At AI Engineer NYC, Anthropic's CJ Avilla and Harrison Stall walked through flows, announced on 9 October: a lead agent writes an orchestration program that starts sub-agents and gathers results, instead of developers configuring a coordinator and roster. In a stock-watchlist demo, Avilla said flows cost "much more" than one agent but found far more detail. Stall said the model can't see its budget mid-run, but developers can set one per session.

On Martin Fowler's site, Specmatic founder Jain had Codex build one API service with and without a Specmatic-built harness of API contracts, examples and tests it runs while coding. Unaided, Luna's service passed acceptance in only 3 of 10 runs; Sol passed all 10 either way, so the harness only raised its cost, about sevenfold. Jain recommends the harness for cheaper models and one final test run for frontier ones, from a single Java service so far.

Modal trained the open-weight Qwen3.6-35B-A3B on coding tasks, generating its practice attempts on GPUs from A100s to B300s, each in its own number format, instead of matching the trainer's hardware. Uncorrected training collapsed, but an off-policy correction, which reweights each token by how far sampler and trainer disagree and drops extremes, matched an all-B200 setup on SWE-bench Pro at about 15% lower sampling cost per update. Modal says any GPU that can serve a model can then contribute samples.

This year Cloudflare's Data team began sending each analytics event across its network once instead of twice, sampling on the servers that collect data and combining matching records into totals at every hop. One in-house analytics product's receivers freed about 750 CPUs and 3.5 TiB of memory. Cloudflare says this halves the data it would otherwise process in five years, though growth erased half of one dataset's saving within two months.

Wing partner Jaipuria points to OpenAI's deployment company, which raised over $4 billion, Anthropic's $1.5 billion Ode and a $1 billion AWS unit, each embedding engineers inside customers to get agents working. On OpenAI's GDPval test of real work tasks, models beat experts on over half, yet most organisations use AI nowhere near that level. With lab models performing similarly on most enterprise work, he argues, whoever embeds the engineers gains a form of lock-in over rival model vendors.

Across five package ecosystems, including npm, PyPI and CRAN, the number of people opening GitHub issues, where users ask questions and report bugs, fell 36% to 64% per repository over the same period, and first-time arrivals mostly fell faster. The decline began years before AI coding tools, which Padgham says very likely accelerated it. His advice to anyone using them: never let one open an unsolicited pull request on a repository you don't know well.

Goci's tool, License Detector, flags dependency licences that conflict with company policy, such as a banned GPL-3.0 package. On his own rerunnable benchmark, it identified licences with 99.0% accuracy at 161 files a second, against 92.8% and 16 files a second for the 2,600-star ScanCode Toolkit. He believes that effort could not have gone into ScanCode instead, and asks how anyone choosing a library will pick between a slow, popular incumbent and hundreds of potentially great but likely unmaintained one-person projects.

OpenCode calls this Code Mode and says it lets you add a thousand MCP servers with minimal context bloat. The new version also runs one background server shared by its terminal, web and desktop apps, and the agent can hand work to background subagents, even on another model, while you keep prompting. Upgrade the CLI from the latest 1.x with one command, then point OpenCode at the 2.0 post to migrate your setup, as the server API changed too.

Beyond text, it returns a box for every line and layout block, labels such as heading or table, and the reading order, as markdown or JSON. Datalab's own benchmark puts it 11-17x faster than Tesseract and PaddleOCR's text models, and ahead of both on text in its default mode. It suits clean printed pages rather than handwriting or complex tables, and installs with pip as a command-line tool and Python library.

Through its MCP tools, an agent can run the app's own commands, change parameters, measure parts, screenshot its work and export the result. GitHub user 0xlillith modelled a rover entirely through the MCP server, 3,323 editable steps that rebuild without errors, though the project says it is early but usable for simple parts. It installs on Windows, macOS and Linux, one command adds it to Claude Code, and you can watch the agent work in the running app.