Gemini 4 Argon launches at half price, and Anthropic replaces Git for its agents
Google ships Gemini 4 Argon at half price for six weeks, a 24B open model targets laptops, and Anthropic explains why Git stopped scaling for agents.
Today’s theme is a quiet one: the model is stopping being the interesting part. Prices, harnesses, hooks and even version control are bending around agents.
1 · Top story
Google’s Gemini 4 Argon is out, and it’s half price for now
Google launched Gemini 4 Argon on Google Cloud yesterday (model ID gemini-4-argon), through the Gemini API, AI Studio and Antigravity. It costs $2 in and $10 out per million tokens until Nov 30, then $4 and $20, with cached input 95% cheaper. Google says it pauses for confirmation before irreversible steps and, on a prompt-injection benchmark from Gray Swan, lets only 0.7% of adaptive attacks through.
Most of the evidence is Google’s own or comes from early-access partners, and the page doesn’t say whether this is GA or preview. Treat the introductory price as a promotion, not a plan: any cost model you build now should be run again at double.
A cheap check: replay 20 of your hardest real agent traces against your current model and weigh cost per resolved task, before the price changes.
2 · Models & features
Liquid AI’s 24B open model is built for agents on your own hardware
LFM2.5-24B-A2B is a mixture-of-experts model with 24B total and 2B active parameters, a 131K context and a reasoning-first design. Liquid AI reports 184 tokens/s on an Apple M5 Max and runtime support for llama.cpp, MLX, vLLM and SGLang. It leads on several tool-use benchmarks but trails Qwen3.6-35B-A3B on SWE-Bench Verified (51.0 against 66.2).
That’s the shape of a good local tool-calling model, not a coding one. The benchmarks are the vendor’s own, and the page doesn’t state licence terms, so read the model card before you plan around it.
3 · Agents
Google’s agent kit now covers five languages, and A2A comes with it
Google’s Agent Development Kit is now available in Python, Go, TypeScript, Java and Kotlin, all with MCP client and A2A support. Go and TypeScript run the new ADK 2.0 graph workflow engine, and Go packages down to a roughly 2 MiB distroless image. Kotlin’s graph workflows are still experimental.
For a team with a Java or Go backend, this removes the usual “rewrite it in Python” step for agent services. The catch is that the post doesn’t label any language as GA or preview, so check the release notes before you commit.
4 · Research
Apple’s Minions pools idle Macs into an inference cloud
Apple researchers describe Minions, a system that turns idle Apple Silicon machines into a pooled virtual GPU cloud for inference, with fault tolerance for devices that come and go. It’s a preprint with no benchmarks on the project page and no code release yet.
The idea is interesting for teams with a fleet of developer Macs and spiky batch inference. The open question is the one that decides whether it’s real: what happens to tail latency when a laptop lid closes mid-request?
5 · People worth following
Geoffrey Huntley’s fix for agents that ignore your rules
Simon Willison quotes Geoffrey Huntley on a coding agent that kept skipping his diagnostics library. Editing AGENTS.md only helped for a while. A pre-commit hook that scans the diff, fails the commit and tells the agent what to use instead got compliance close to perfect.
I think that’s the right instinct: instructions are advice, hooks are policy. His broader advice is to delete AGENTS.md and skill files regularly and test whether they still help. It’s one practitioner’s report, not a study.
Pick one rule your agents keep breaking, and count how often the violation recurs over a week of agent PRs.
6 · Architecture
Anthropic replaced Git because its clients are mostly agents now
Anthropic’s engineering post describes Scion, an in-house source control system built after merges grew from about 2,500 a day in May to over 30,000 by September. The monorepo’s history is nearly 100 GB and CI does roughly 25 million checkouts a day. Scion assumes clients are always online and does history, blame and merges on the server.
Anthropic says it cut CPU and memory 200x against the old mirror fleet and took checkouts from minutes to seconds. It also dropped offline support and global merge locks, trading them for post-merge testing. Anthropic isn’t releasing Scion, only sharing the design ideas, so the transferable lesson is to check which of your tools were designed around a human sitting at a laptop.
That is enough for one morning. The price tag can wait until you have a trace to replay.
Middle East, India & events · 1
- Sarvam’s Bulbul V4 Flash, a text-to-speech model for 11 Indian languages — Sarvam says time-to-first-audio is under 250 ms in production, up to 40% faster than V3, at ₹20 per 10,000 characters against ₹30 before. It is available through the Sarvam API, Playground and Content Studio, and existing Bulbul users switch on the same API. The latency figures are the company’s own. Source
Researched and drafted with AI, checked against the linked sources. How we use AI →