Image: Google (official) — Gemini 4 Argon blog header from the Google Keyword blog (Oct 13, 2026)
Gemini 4 Argon is Google’s newest frontier model, rolled out globally on October 13, 2026. The marketing pitch is “hardest problems”; the practical question is simpler: where do you turn it on, what plan do you need, and which knobs actually change the result? This guide sticks to Google’s own AI Studio pages and Gemini API docs—how to try Argon without code, how to call gemini-4-argon from Python or JavaScript, how thinking levels and long decode work, and what breaks if you migrate from an older model.
1. Check that you can actually use it
Availability depends on the surface, not just “having Google AI”:
- Google AI Studio: paid Gemini API tier, or a Google AI Ultra plan. Free-tier API access is not enough for Argon.
- Gemini app and AI Mode in Search: Google AI Ultra or Pro subscribers.
- Google Antigravity: Google AI Ultra (and enterprise paths Google lists separately).
- Gemini API pricing: introductory rates through November 30, 2026—$2 per 1M input tokens and $10 per 1M output tokens (including thinking tokens). After that window Google lists $4 / $20. Cached input via implicit caching is discounted 95% on hits. Argon is not on the Free tier or Google AI Pro for API use.
If your menu never shows Argon, the usual cause is plan or billing, not a missing toggle.
2. Try it in AI Studio before you write any SDK code
Google’s Build-with-Argon page is aimed at people who want a working preview first. Three low-friction paths:
- Open Google AI Studio, pick gemini-4-argon in the model menu, and run a starter prompt in the Playground (orbital sandbox, photo-to-3D scene, or a financial simulator with live sliders).
- Open the Argon app showcase, click Try on a launch app to play it full screen, or click Remix to open a copy in AI Studio Build and change style, data, or features in plain English.
- In Build, describe the app—or upload a sketch or screenshot—and let Argon write and update the code while you watch the live browser preview. You do not need to start from a blank repo.
My habit: remix one showcase app that is close to what you want, then ask for one change at a time. That keeps the conversation short and makes it obvious when Argon drifted from your brief.
3. Get an API key the right way
For production calls you need a Gemini API key tied to a Google Cloud project with billing. In AI Studio:
- Open the Dashboard → Projects. Import an existing Cloud project if AI Studio does not already show one.
- Go to API Keys and create a key. Keys created in AI Studio since May 28, 2026 are auth keys by default (bound to a service account and restricted to the Gemini API).
- Set
GEMINI_API_KEY(orGOOGLE_API_KEY) in your environment. The official client libraries pick these up automatically; if both are set,GOOGLE_API_KEYwins.
Do not paste keys into front-end apps. Google’s own security notes are blunt: unrestricted standard keys are rejected, and dormant unrestricted keys can be blocked. Restrict keys to the Gemini API and keep them out of git.
4. Your first Interactions API call
Argon’s model ID is gemini-4-argon. Default thinking level is medium, and a normal (non-background) response caps around 64K output tokens unless you opt into long decode. Minimal Python:
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-argon",
input="Write a three.js script that renders a realistic 3D black hole with gravitational lensing.",
generation_config={
"thinking_level": "medium" # low | medium | high
},
)
print(interaction.output_text)
The same shape works in the JavaScript @google/genai client and via REST on /v1beta/interactions with header x-goog-api-key.
5. Pick a thinking level on purpose
Thinking level is the main quality / cost / latency dial:
- low — fewest tokens, fastest first token; fine for simple or cost-sensitive jobs, weaker on hard tasks.
- medium (default) — Google’s recommendation for most complex code and agent work.
- high — deepest reasoning; pairs with long decode for long proofs, big refactors, or heavy science.
Run the same eval at all three levels and keep the lowest that still passes. Raising thinking without raising max_output_tokens can truncate the answer, because thinking and output share the same limit.
6. Long decode when 64K is not enough
Argon supports a unified 1M-token window shared across input, thinking, and output. To go past ~64K tokens of output on the Interactions API:
- Set
background=True. - Set
max_output_tokensgreater than 65,536 (examples in the docs use 131,072 or 262,144). - Poll
interactions.getuntil status leavesin_progress/queued. - Join text from every
model_outputstep—each continuation hop adds another step.
If you enable background without raising max_output_tokens, the server still stops at 64K. For hard prompts, Google also advises HTTP client timeouts of at least 15 minutes, or relying on background execution so your laptop does not have to stay open for every hop.
7. Migration gotchas from older Gemini models
Google’s migration checklist is short and easy to miss:
- Change the model string to
gemini-4-argon. - If you set a low
max_output_tokenstoday, raise it—at least 32,768 is recommended because Argon spends more of the budget on thinking. - Remove explicit context caching (
cachedContents/caches.create); it is not supported. Implicit caching still applies for prompts of at least 4,096 tokens. - Drop
temperature,top_p, andtop_k—they are ignored if sent. - Keep multi-turn state with
previous_interaction_idinstead of resending the whole transcript when you can.
If you use a coding agent with skills, Google documents a gemini-api-dev skill that can automate much of that migration.
8. A practical first afternoon
- Confirm Argon appears in AI Studio on a paid / Ultra path.
- Remix one showcase app and change only the visual style or sample data.
- Create a restricted API key, set
GEMINI_API_KEY, and run the black-hole (or your own) quickstart atthinking_level: medium. - Repeat the same prompt at
lowandhigh; keep whichever meets your quality bar at the lowest cost. - Only then try one background long-decode job on a real hard prompt, with a high
max_output_tokensand a billing alert already set in Cloud.
Honest limits
Argon is expensive relative to flash-class models, and thinking tokens count as output. It is unavailable on free API access. Long decode and high thinking can take minutes. Agent features (Antigravity / Managed Agents) need the matching subscription and, for some orgs, admin enablement. Treat the first week as supervised: check numbers, permissions, and irreversible actions yourself.
Sources
- Google Blog: Gemini 4 Argon is now available for everyone (Oct 13, 2026)
- Google AI Studio: Build with Gemini 4 Argon
- Google AI Studio: Gemini 4 Argon developer guide
- Gemini API docs: What’s new in Gemini 4 Argon
- Gemini API docs: Long decode
- Gemini API docs: Using Gemini API keys
댓글
댓글 쓰기