Argon goes GA, two open giants drop, and Claude Code stops burning your cache

Welcome to the first Vibe Diff. Every Monday you get what changed in the tools you ship with, why it matters, and one prompt you can copy. This one covers Oct 12–19, 2026.

1. Gemini 4 Argon is GA, and it's cheap until Nov 30

Google made Gemini 4 Argon generally available on Oct 13. The intro price is $2 input / $10 output per million tokens, and cached input is 95% off. After Nov 30, 2026 it goes to $4 / $20. You get a 1M unified input+output token limit, and it's live in AI Studio, the paid tier of the Gemini API, and Antigravity (Ultra). Google says it ranks #1 on the Vals Index, and Arena reports Gemini 4 Argon (High) at #5 in Agent Arena.

Why it matters to builders: you have about six weeks at half price to find out whether Argon beats your current default on your own code. Heavy prompt caching makes the trial even cheaper. Just don't lock a budget to the intro price.

2. Meta's Muse Spark Open: big, multimodal, open weights

Meta released Muse Spark Open on Oct 15. It's 816B total / 42B active parameters, has a 1M context, and is natively multimodal. Meta says it runs on two Apple M5 Ultra Mac Studios with MLX. Hosted APIs are available through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together, and the weights are on Hugging Face.

Why it matters to builders: you can now call a frontier-class open multimodal model from seven providers, so it's easy to try without self-hosting. Read the Muse Model License before you ship. Meta describes its commercial terms as limited.

3. Reflection Beam 1.0: Apache 2.0 and strong at code

Reflection released Beam 1.0 on Oct 16 (announcement). It's a 501B MoE with 23B active parameters, a 1M context, and text only. The license is Apache 2.0. It reports 81.6% on SWE-bench Verified (BF16) and 80.1% on Terminal Bench v2.1. Checkpoints come in BF16, FP8, and NVFP4, it ships with a Rust harness, and transformers, vLLM, and SGLang support it from day one. On Hugging Face Inference Providers you can call it through Baseten, Fireworks, and Together. There's also $10,000 in Hub credits for the best projects built on it.

Why it matters to builders: a permissive license plus coding-agent-grade scores makes this the open model to try for agent backends you want to own. Self-hosting is still a lot of hardware, though: the NVFP4 checkpoint alone is about 280GB.

4. Claude Code stops letting your prompt cache die

The Claude Code releases from v2.1.299 to v2.1.302 (Oct 14–16) added idle compaction. If a conversation over 200K tokens sits idle for about 54 minutes, Claude Code compacts it before the 1-hour prompt cache expires. You can tune the threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS or turn it off with "idleCompaction": false. v2.1.302 adds a toggle for it in /config and a CLAUDE_CODE_AUTO_MODE_MODEL variable. It also turns on fork subagents by default in claude -p, the SDK, desktop, and VS Code, and warns that standalone Claude Design (claude.ai/design) closes by December 14, 2026. v2.1.301 added CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS.

Why it matters to builders: long sessions that you leave for lunch get cheaper to pick back up. If you script claude -p in CI, check how the default fork-subagent change affects your token use and output.

5. Cursor opens the Origin App Marketplace

Cursor's changelog for Oct 13 announced the Origin App Marketplace with 20+ apps, including Cloudflare, Netlify, Render, Vercel, CircleCI, CodeRabbit, Greptile, Semgrep, Socket, Sentry, and Linear. Anyone can submit an app.

Why it matters to builders: deploys, code review, security scanning, and issue tracking can now plug into the agent directly, without you pasting context in by hand. If you build dev tools, this is a new place for people to find you.

6. Lovable lets browser agents use your app

Lovable's changelog adds "Let browser agents use your app". It exposes WebMCP tools that run in the user's own tab. By default only read-only or easy-to-undo tools are exposed, and destructive tools are added only when you ask for them. The same update adds Gemini 4 Argon for in-app AI features, an "Always allow all tools from this connection" option for connectors, and a change so Lovable Cloud refuses unsupported CREATE EXTENSION statements.

Why it matters to builders: the next "user" of your app might be an agent. Starting with safe, undoable tools is a good default to copy, even outside Lovable.

7. Replit: Bring Your Own Cloud, Haiku 5.5, and a partner directory

Replit's Oct 16 changelog adds Claude Haiku 5.5 to the Power Mode model selector for Core and Pro. It also announces Bring Your Own Cloud: Full BYOC is in private preview, while BYOC for Published Apps and Image Exports are in early access. Separately, Replit launched a Solutions Partners directory (Oct 13).

Why it matters to builders: BYOC answers the usual objection that a client's data can't leave their cloud. If you sell Replit-built apps to companies, this could unblock deals.

Copy this: a 30-minute model bake-off on your own repo

Benchmarks are someone else's code. Before you switch your default model (Argon, Beam, or anything else), run the same task through each candidate and compare.

You are evaluating yourself as a coding agent on this repository.

Task: <paste one real, recently closed issue or bug from this repo, with the original description>

Rules:
1. First, list the files you need to read and why. Read them before editing.
2. Write a short plan (max 6 bullets) before any code.
3. Make the smallest change that fixes the issue. No unrelated refactors.
4. Add or update a test that fails before your change and passes after.
5. Finish with: files changed, commands to verify, and anything you're unsure about.

How to score it: run the prompt on each model in a fresh branch. For each run, note pass/fail on the test, diff size, wall-clock time, and cost (from your provider dashboard). Do three different issues, not one. Pick whichever model passes most often, and use cost to break ties.

Tool of the week: Vercel Token Pass

Vercel Token Pass (Oct 15, early access for Pro teams) costs $30/month and gives you $80 of monthly credit for GLM 5.3 and Kimi K3 through AI Gateway. It works with fx, Command Code, Pi, or the AI SDK. Zero data retention is on by default. Unused credit doesn't roll over, and requests stop when the credit runs out, so there's no overage bill.

Why we picked it: you get a hard spending cap and a cheap way to put two strong open-weight coding models into the workflow you already have. That's a good fit for side projects and for running the bake-off above.

---

That's the diff. Reply and tell me what you shipped this week. I read everything.