tool notes · from issue #1
Gemini 4 Argon: pricing, who it's for, and our take
· prices verified against Google · covered in issue #1
Google's frontier model for long-horizon coding and hard multi-step work — with intro API pricing through Nov 30. Here's the numbers, who should use it, and how we slot it next to a cheap grunt model.
What it is
Gemini 4 Argon (released 13 Oct 2026) is Google's flagship reasoning model for complex software engineering, long refactors, and autonomous multi-step agents. Google's launch post cites a Chromium refactor that untangled a ~6M-line legacy folder, plus use in Antigravity for agentic coding. The Gemini API gives it a 1M unified input-output token limit.
API model id: gemini-4-argon. Paid Gemini API tier only (not Free tier, not Google AI Pro). Also in Google AI Studio / Antigravity (Ultra), the Gemini app and AI Mode in Search (Ultra/Pro), and GitHub Copilot for Pro+, Max, Business, and Enterprise (usage billed at provider list pricing). See Google's launch post and Gemini API pricing.
Pricing verified 16 Oct 2026
Standard paid tier, USD per 1M tokens. Output price includes thinking tokens. Implicit context caching only (explicit caching not supported):
| Period | Input | Output | Cached input |
|---|---|---|---|
| Through 30 Nov 2026 | $2.00 | $10.00 | $0.10 |
| From 1 Dec 2026 | $4.00 | $20.00 | $0.20 |
Batch API: 50% off input and output ($1/$5 intro, then $2/$10). Sources: Google launch, ai.google.dev pricing.
The intro window is the search-worthy part: same model at half the list price until Nov 30. After that it sits at Opus-class $4/$20 unless you use Batch or cache-heavy shapes.
Who it's for
Use it when…
- Big, multi-file jobs: migrations, refactors, architectural untangles across a real codebase
- You want a frontier planner that can hold a long PLAN.md and keep working across steps
- You're already on Copilot Pro+ (or Max/Business/Enterprise) and want Argon in the model picker
- You can burn paid API / Ultra quota and will finish the hard path before the intro price ends
Skip it when…
- The job is high-volume grunt work (summaries, call-site sweeps, classification) — use a cheap subagent like Haiku 5.5
- You only have Free-tier Gemini API or Google AI Pro — Argon isn't on those
- You need one model to do everything cheaply; Argon is the expensive "plan big" slot, not the default loop
Our take
We covered Argon in Commit Weekly #1. The short version: yes, for the big jobs — not for every turn.
The workflow we actually recommend is plan big / grunt small / prove it:
- Plan with a frontier model — Argon is the "plan big" / long-horizon slot: have it write PLAN.md (files, risks, definition of done) and wait for your OK.
- Hand the grunt work to a cheap model — summaries, "find every caller of X," extract-this-line jobs go to something like Haiku 5.5 so you don't torch the Argon budget.
- Prove it at runtime — run real flows against the real stack; the test run says "done," not the agent.
Argon is step one. Expensive enough that you shouldn't burn it on lookups; capable enough that the plan is worth the spend — especially while intro pricing lasts.
Read the full issue write-up on issue #1, the companion Haiku 5.5 notes, or browse the archive.
// subscribe
Get tools like this every week free
One email a week with the AI tools worth trying. Subscribing happens on Beehiiv, our newsletter platform.
No form? or subscribe on Beehiiv → · free · Tuesdays · unsubscribe anytime
← Home · Issue #1 · Haiku 5.5 · Sources: Google launch, Gemini API pricing, Copilot changelog