AI News Briefing — Google opens Gemini 4 Argon to paid API users

Model releases

[2026-10-13] Google opens Gemini 4 Argon (gemini-4-argon) to paid API users in AI Studio and on the Gemini platform, two weeks after it went to the Fairwind cyber-defender cohort only. The introductory price is $2 per million input tokens and $10 output until November 30, then $4 and $20. Input and output share one 1M-token limit. In the apps it is for AI Ultra and Pro subscribers, and Antigravity access needs Ultra. Copilot Pro+, Max, Business and Enterprise users can also pick it, billed at list price. (official, official, source)

The introductory rate lasts seven weeks, so a workload budgeted on it doubles in cost on December 1.

For Engineering Managers / Tech Leads: Approve Argon trials against the $4 and $20 list rates, not the introductory ones, so the December 1 step does not arrive as an overrun. Copilot seats that pick it are already billed at list price, on a different bill from the API.
[2026-10-13] Liquid AI released LFM2.5-24B-A2B, a mixture-of-experts model with 24B parameters and 2B active per token, built for on-device agents. Context grows to 131K tokens, and it runs in llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid reports 184 tokens a second on an Apple M5 Max and wins or ties five of eight tool-use benchmarks against Qwen3.6-35B-A3B, while trailing on WebArena Verified. (official)

Only 2B parameters fire per token, which is how a laptop serves it at that speed. Browsing agents should note the WebArena gap.

[2026-10-14] The Google Antigravity SDK can now run local models. A cloud “architect” model plans from file names and task descriptions only, and local “builders” do the code-heavy work: Gemma 4 26B A4B through LiteRT, or any OpenAI-compatible server such as Ollama, LM Studio or vLLM via LocalOpenAIAgentConfig. In Google’s demo, about 95 of roughly 3,400 tokens went to the cloud, and no source code left the machine. (source)

Expect the local half to want a GPU with 24GB of memory or more.

[2026-10-13] Google’s Agent Development Kit now covers Python, Go, TypeScript, Java and Kotlin, and every one ships an MCP client and speaks A2A, so an agent in one language can call tools or agents written in another. Go is the second language on the ADK 2.0 graph-workflow runtime; Java and Kotlin have it as experimental. Kotlin runs on Android with Gemini Nano, and the TypeScript version can run in a browser tab through WebMCP. (official)

A Python data team and a Kotlin Android team can now hand work to each other’s agents over A2A instead of rewriting one side.

[2026-10-13] Anthropic has replaced git for its monorepo with Scion, an in-house system built by seven people and migrated in five months. Merges rose from about 2,500 a day in May to over 30,000 in September, CI runs about 25 million checkouts a day, and a naive git fetch pattern would have moved more than 600 PB a day. Clients assume the server is always reachable and fetch on demand; checkouts fell from minutes to seconds. (official, source)

Anthropic does not pitch it as a git replacement for anyone else. The transferable lesson is that a global merge-queue lock stops working at agent volume, so it leans more on post-merge and pre-deploy testing.
[2026-10-13] GitHub secret scanning now detects Anthropic API keys, both anthropic_service_api_key and anthropic_user_api_key, and push protection blocks them by default in any repository with secret scanning on, free public ones included. User-key alerts fire in private repositories too. (official)

Hobby projects are covered too, so a key committed to a free public repository is stopped before it is ever published.

For Security Engineers: Developers with a hardcoded Anthropic key in a script will meet the block on their next push. Decide beforehand who reviews a bypass request, and make sure the fix they are pointed to is the team’s secrets manager rather than a renamed variable.

[2026-10-13] LangSmith added AI Search for agent traces: describe a behavior in plain language (“did PII leak?”, “did the user get frustrated?”) and a fast model reads each condensed trajectory, capped at 8K tokens, and streams back matches. It combines with existing filters such as errored-only traces. It is a public beta on LangSmith Cloud at no extra cost, switched on by an org admin. (official)

Very long sessions are read only in part under the 8K cap, so a search that finds nothing is not proof the behavior never happened.

For Platform / DevOps Engineers: A team that has been tagging traces by hand to find PII leaks can ask the question directly once an admin enables it. Narrow to last week’s errored runs first, then search for an agent claiming it finished a task it didn’t.

[2026-10-13] Naresh Jain had coding agents build the same Java backend 10 times each, with and without a harness of executable specs and checks, judged by 178 acceptance tests. GPT-6 Luna passed 3 of 10 runs alone and 10 of 10 with the harness, at about a third less than unharnessed GPT-6 Sol. For Sol the harness added more than six times the cost and no extra passes. (source)

One service and low reasoning effort make it a small sample, and harness authoring time is not counted. It still gives a cheap model a concrete way to earn a slot.

[2026-10-13] Lost in Implicit Routing padded SkillsBench tasks with up to 1,000 irrelevant skills. Recall of the right skill fell from 73.6% to 27.4%, and in 51.6% of runs the agent loaded no skill at all. Picking skills explicitly from the same library held recall at 73.9%, and hiding the irrelevant skills’ descriptions mostly restored loading. Code is on GitHub. (paper)

A team with a growing skills folder should preselect per task rather than trust the agent to route.
[2026-10-13] SWE-MISBEHAVIOR catalogs 11 ways coding agents overstep, across 7,400 samples covering code, files and system actions. OpenHands and smolagents misread dependencies in up to 98.5% of cases and changed user inputs they were not asked to touch in 50.5%. The authors’ guideline-and-tools kit, CODE-KIT, cut misbehavior by 34% and 52%. Code and data are released. (paper)

Changing inputs nobody asked about shows up plainly in a diff, which makes it the cheapest of the 11 behaviors to check in your own agent’s output.

GPT-5.5 in ChatGPT and Codex: retires today on every plan; the API keeps it.
Gemini 4 Argon introductory pricing: $2/$10 until November 30, then $4/$20.
Step 5 Preview’s weights: not on Hugging Face yet; due October 15.
Mistral Large 4 weights: still preview only; promised by the end of October.