Google releases Gemini 4 Argon at half its list price until November 30, Meta launches Muse Wonder, GitHub blocks Anthropic keys on push
Two new frontier models in two days, and the more useful details are around them: Argon is priced to be tried now and paid for later, Copilot switched it on by default, and GitHub and Copilot added controls that fail closed. New models are cheap to test this month; the controls around them are what to put in place first.
Introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input 95% off; after November 30, 2026 it becomes $4 and $20.
Google says Argon ranks first on the Vals Index and that inside Google it raised agentic code submissions in Antigravity by more than 35% while cutting rollbacks by nearly a third.
Examples include a 2.85x end-to-end vLLM speedup for Gemma 4 E2B on an A100 (about 37,900 to over 108,000 tokens per second) and a 40% cut in in-memory session prep time for the open-source ADK.
Shortcut attempts on coding benchmarks dropped to 2%, and attack success on the Gray Swan indirect prompt injection benchmark fell to 0.7% across 15 adaptive attempts.
Builder's takeThe introductory price is the real news: for seven weeks a frontier model costs half its list price, so this is the cheapest window to run it against our own evals. I’d put Argon through the scoring prompts behind the AI interview product now, but write the comparison at the $4/$20 list price, because that is what production will pay after November 30.
Available to Copilot Pro+, Max, Business and Enterprise users in VS Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, JetBrains IDEs, Xcode and Eclipse; rollout is gradual.
GitHub says that in early testing Argon solved more tasks than previous Gemini models on every coding benchmark, with its largest gains in Copilot CLI.
It is billed at provider list pricing under usage-based billing, and new models are enabled automatically unless an administrator has turned off the global default or disabled this model.
Builder's take“On by default, billed at list price” is the line team leads should read. If you run Copilot Business, decide this week whether Argon goes on the allow list, and set a budget before people find it in the model picker. Pair it with the 10-10 change that lets code review be billed to the organization: model choice is now a cost line, not a preference.
Muse Wonder is available today to developers through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers.
Both models have “Active Memory”, carrying relevant details across conversations, and “Active Focus”, retrying other approaches, recovering from errors and checking in before anything important.
Meta says the models are trained not to share private information without permission, and that a full report with additional safety evaluations will be available soon.
Builder's takeMeta is now a frontier API vendor, but the announcement gives no price and no benchmark numbers, and the safety report isn’t out yet. I’d register for the API and try one long agent task on it, but hold off on routing user data to it until the safety report and pricing are public; cross-conversation memory is exactly the feature that needs a clear data policy first.
24B total parameters with 2B active per token; the context window grows from 32K to 131K tokens and pre-training from about 17T to 33T tokens.
It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B and 27.69 on AutomationBench versus 8.72, but trails it on SWE-Bench Verified (51.00 vs 66.20) and WebArena Verified (26.75 vs 37.62).
Agentic RL was run inside OpenClaw, Hermes Agent, Pi and OpenCode; day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX.
Both the instruct and base models are open-weight on Hugging Face.
Builder's takeRead the table, not the headline: it wins on tool calling and instruction following and loses clearly on coding and web tasks. That makes it a candidate for the local tool-routing and form-filling steps in PandaClaws, not for a coding agent. Given the 10-13 self-replication research, if you self-host it, keep the endpoint off the public internet.
New detectors: anthropic_service_api_key, anthropic_user_api_key and cloudsmith_api_key; Salesforce joins as a secret scanning partner for Marketing Cloud OAuth2 tokens.
Repositories with secret scanning enabled, including free public ones, will have commits containing the two Anthropic key types blocked automatically.
In a companion post the same day, GitHub said its fine-tuned ModernBERT secret classifier cut median post-push scan time by 86% and serving cost by more than 98% versus its previous LLM pipeline.
Builder's takeThis closes one of the leaks the 10-13 self-replication research pointed to: stolen API keys paying for someone else’s inference. A blocked push is still a key that sat in a working tree, so I’d keep model keys in a secret manager and rotate any that ever touched a repo. The classifier numbers are also a useful pattern: a small fine-tuned model replaced an LLM call at under 2% of the cost.
Setting forceRemoteSettingsRefresh to true makes Copilot CLI, VS Code and the Copilot app fetch current server-managed settings at startup; if the client cannot confirm current policy, it fails closed.
The setting can be delivered through server-managed, MDM-managed or file-based settings; MDM and file delivery can require freshness even before first sign-in.
When the sandbox key is in managed settings, users pinned to older Copilot clients must update to keep using Copilot.
Builder's takeThis is the same move as Claude Code’s fail-closed hooks on 10-10: when the control can’t be confirmed, the agent doesn’t run. I’d turn it on only for the policies that matter, such as sandbox and network rules, and test it on a laptop with flaky VPN first, since fail closed also means nobody codes when the policy server is down.
telemetry.capture.identity adds user.name (the GitHub login), process.user.name (the local OS account) and host.name to OTel data; it is off by default.
Prompts, responses, tool arguments, tool output and policy detail can each be captured independently, and all five are off by default.
Explicit managed values take precedence over local settings, SDK configuration and environment variables.
Builder's takeSeparating identity from content is the right default: you can answer “whose agent ran this command” in an incident without storing everyone’s prompts. I’d enable identity plus tool arguments for audit, keep prompts and responses off, and apply the same split to the agent logs in our own products.
Grok 4.7 is served through two cross-Region inference profiles, us.xai.grok-4.7 for US data residency and global.xai.grok-4.7, and supports OpenAI-compatible Chat Completions and Responses APIs plus the Bedrock Converse API.
Across three search-style tasks run six times at each effort level, the agent answered 67 of 72 runs correctly; xhigh got all 18 right.
Citing SpaceXAI, the post lists a 500K-token context window and Terminal-Bench 4.0 rising from 20.3% on Grok 4.6 to 37.6%.
Builder's takeThe method matters more than the model: run your task several times at each effort level and price it in tokens, then pick the lowest level that stays correct. That is the same lesson as GPT-6.1 Sol’s Ultrafast tier on 10-11, from the other direction. If you already run on AWS, Bedrock now lists GPT-6 and Grok 4.7 side by side, so an A/B on your own prompts is cheap to set up.
RTX Spark laptops, with up to 128GB of unified memory and 1 petaflop of AI compute, go on sale Friday, October 16.
GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy are updated for RTX Spark, with security through OpenShell and Microsoft Execution Containers (MXC) and hybrid local and cloud routing.
NVIDIA Nemotron 3.5 Bolt 74B, built for agentic AI on RTX Spark, releases Thursday, October 15; NVIDIA PAIR leaves beta and adds llama.cpp support.
llama.cpp adds support for decision models and releases a Windows app.
Builder's takeTogether with LFM2.5 today, the local agent stack is filling in: hardware, small models, harnesses and llama.cpp support for decision models. I’d plan for “hybrid local and cloud routing” as a product feature, starting with keeping file parsing in the AI cloud drive on the user’s machine when it can, and only sending the hard questions out.
Researched and drafted with AI assistance; editorial standards and views set by Darius.