Sig Weekly

Gemini 4 Argon's price is set to double Nov 30. Plus an Apache 2.0 coding model

Two big open-weight coding models came out of US labs this week, and only one of them ships under Apache 2.0. To me that matters more than any benchmark chart this week. If you build agent products, the license decides whether you can put the model inside your product without calling a lawyer. The other date to remember is November 30. That's when Gemini 4 Argon's launch price is reported to double. If you plan to test it, do it this month. I haven't run these models on our own work yet, so below is what the labs say, plus my read on it.

Written by Mara at Sig · Oct 19, 2026

Two big open-weight coding models came out of US labs this week, and only one of them ships under Apache 2.0. To me that matters more than any benchmark chart this week. If you build agent products, the license decides whether you can put the model inside your product without calling a lawyer.

The other date to remember is November 30. That's when Gemini 4 Argon's launch price is reported to double. If you plan to test it, do it this month.

I haven't run these models on our own work yet, so below is what the labs say, plus my read on it.

The week in AI

Google opens Gemini 4 Argon to developers at an introductory $2/$10 per million tokens

Google's new frontier model is live in the paid Gemini API, AI Studio and Antigravity, aimed at long, multi-step work like large refactors, finance, legal and security tasks.

Why it matters: The benchmark claims are Google's own, so I'd discount them until independent numbers land. The pricing is the part I trust: $2 in and $10 out per million tokens today, reported to go to $4 and $20 on Nov 30. Cached input at 95% off is what makes it interesting for agents that reread the same repo all day.

What to do: Run your hardest real refactor through gemini-4-argon this month and write down the cost per task. Decide before Nov 30.

Source: Google: Gemini 4 Argon is now available for developers

Reflection ships Beam, a 501B open-weight coding model under Apache 2.0

Beam has 23B active parameters, comes in BF16, FP8 and NVFP4, and launches with an open-source terminal agent called Mirror CLI.

Why it matters: This is the one I'd watch. Apache 2.0, 23B active parameters, and it ships with its own terminal agent. Jack Clark's Import AI points out that Beam trails the best Chinese open models on several benchmarks, so it's probably not the strongest open model. It may be the easiest one to ship inside a commercial product.

What to do: If you pay for a hosted coding model, run Beam side by side on five of your real tasks. Reflection's API is free to start in public beta.

Source: Reflection: Beam is now available

Meta releases Muse Spark Open, an 816B open-weight model that runs on two Mac Studios

The mixture-of-experts model has about 42B active parameters, a 1M-token context window and takes text, images, audio and video. Meta says it is the most capable US open-weight model to date.

Why it matters: Running a model this size on two Mac Studios is a real step for self-hosting. The catch is the Muse Model License, which Meta itself describes as having limited commercial terms. That's the line most people will skip.

What to do: Read the license before you plan a product around it. To just try it, use a hosted provider like OpenRouter or Together first.

Source: Meta AI Research: Introducing Muse Spark Open

Claude Code now compacts idle conversations, and Codex turns Daybreak back on

A week of coding agent changes: automatic idle compaction for long Claude Code sessions, new environment variables, and a Codex default that flipped twice in a week.

Why it matters: Idle compaction saves money on long sessions. It also swaps your conversation for a summary while you're away, and summaries drop details. I'd rather choose when that happens.

What to do: If you leave long sessions open, check /config after you update. Set "idleCompaction": false if your work depends on exact earlier context.

Source: DEV Community: This week in Claude Code, Codex and Gemini CLI (week of Oct 18, 2026)

Tool worth trying

mcpkit. If you build an MCP server, this checks it against 12 clients and tells you what breaks in each one, with a source for every finding. Open source, MIT license.

npx @vrajved/mcpkit --help

01 · Agent-built apps
cafecito
A coffee shop your AI agent orders from
02 · Dev tools
Guilds
Open-source AI agent teams you can inspect and debug
03 · Dev tools
Brocoded
Approve Claude Code and Codex actions from your phone
04 · Dev tools
Agentboxd
Real inboxes for AI agents, with prompt-injection checks
05 · Agent-built apps
Tile Tunes
A piano tiles game vibe-coded with Claude Code
06 · Agent-built apps
ilovecollage
A vibe-coded collage and meme studio in the browser
07 · Agent-built apps
3rd Mainland Bridge
A vibe-coded game: cross Lagos' 3rd Mainland Bridge on a bike
08 · Dev tools
Cakie
Turn screen recordings into prompts for coding agents
09 · Agent-built apps
Vball Manager
A volleyball management sim built with Claude Code
10 · Agent-built apps
CyberFlow
A security score for vibe-coded apps, built solo with AI agents

Get your referral link: subscribe free and it comes in every issue. Refer 1 friend and you get the whole Sig blueprint archive. Subscribe.

Sig Weekly

Get the next issue on Monday.

Back to all issues