Skip to content

Issue #1

Two frontier models, one $0.10 model, and a QA bot for your agent

· 5 min read

Welcome to Commit Weekly #1. Google and Meta both shipped frontier coding models, Anthropic made its small model really cheap, and "prove it works" got its own tool. Here's what's worth your time.

This week's 5

1. Gemini 4 Argon

What it is: Google's new frontier model for long-horizon work like big refactors and migrations, with a 1M input-output token limit in the API.
For: Builders with a gnarly codebase job they keep putting off.
Worth trying? Yes, for the big jobs. Google says it helped untangle a ~6M-line legacy folder in Chromium. It's already in GitHub Copilot (Pro+ and up).
Price: $2 / $10 per 1M input/output tokens until Nov 30, then $4 / $20. Paid API tier only.

2. Muse Wonder

What it is: Meta's first frontier coding models. Wonder is on the Meta Model API and in Muse Code. Wonder Full is subscription-only.
For: Builders who want another frontier option.
Worth trying? Kick the tires. Meta says Full tops DeepSWE and SWE-Marathon, but those are Meta's numbers. Test it on your own repo before you switch.
Price: $4 / $20 per 1M tokens ($0.20 cached input), 1M context. Full needs a Power or Maximum Usage plan (prices weren't in the post).

3. Claude Haiku 5.5

What it is: Anthropic's fast, cheap small model, now with an adjustable effort setting.
For: Multi-agent setups where grunt work eats the budget.
Worth trying? Yes, as a subagent. Even Anthropic says Sonnet and Opus are still better for complex agentic coding.
Price: $0.10 / $0.50 per 1M tokens for prompts up to 100K, and $0.50 / $2.50 above that.

4. Kerno v1

What it is: A runtime QA agent for Claude Code, Cursor, Codex or any MCP agent. It tests your backend flows against your real DB and services, then re-runs them on every PR.
For: Vibe coders whose backend "works" until somebody actually logs in.
Worth trying? Yes, if agents write your backend. It covers what code review and mocks miss. Setup takes 30–60 minutes.
Price: Free Community plan (1 repo, 1 user, 30 runs/month). Pro is $50 per dev per month.

5. Vani Code Agent

What it is: A free VS Code agent extension. Edits come as diffs, commands need your approval, and .env and .git are protected.
For: Cautious beginners who want an agent that asks first.
Worth trying? Maybe. The defaults are sensible, but it's in preview and tied to the Vani API.
Price: The extension is free. Usage is billed through your Vani API key.

Steal this workflow: "Plan big, grunt small, prove it" (ours)

  1. Plan with a frontier model. Prompt: "Read this repo. Before writing code, write PLAN.md: the files you'll touch, the risks, and how we'll know it works. Stop and wait for my OK."
  2. Hand the grunt work to a cheap model. Send summaries and "find every caller of X" jobs to a small subagent like Haiku 5.5.
  3. Prove it at runtime. Prompt: "Run the affected user flows against the real local stack (not mocks) and paste the evidence. If anything fails, fix it and run again. Don't tell me it's done until it passes."
  4. Commit the proof. Keep the tests in the repo as your baseline.

The agent doesn't get to say "done." The test run does.

Quick hits

That's the commit. Ship something this week.

— Commit Weekly

Forwarded this? Subscribe free.

← All issues