Why Gemini 4 Argon is the big model in Gemma Gemini

Since October 13, 2026, the hard messages on Gemma Gemini have gone to Gemini 4 Argon, through OpenRouter as google/gemini-4-argon with reasoning effort set to high. Everyday questions stay with Gemma 4 E2B on NVIDIA Jetson Orin Nano boards in Offline Base's fleet.

From Gemini 3.8 Flash to Argon

The first version of Gemma Gemini, built on October 3, used Gemini 3.8 Flash. Google's Gemini API release notes called it "our most intelligent Flash model" when it became generally available on September 2. The same notes mark Gemini 4 Argon generally available on October 13 and call it "our most capable model built for complex coding, agentic, math, and science tasks." Gemini 4 Argon: what Google released covers the launch.

We switched that day because of what the cloud side handles. Offline Base's position is that small open-weight models can handle most everyday questions and that frontier models are tools to call on when needed. In Gemma Gemini, the cloud model mostly gets messages where the router gives Gemma less than an even chance: proofs, code, and health, legal and money questions. It also gets anything with a file. A cloud side that sees mostly hard questions should run the model Google built for them.

The Vals Index, an outside benchmark of agentic finance, coding, legal and tax tasks, ranked Argon first at 68.90% when it was updated on October 13. Several of those fields, such as code, legal and money questions, are also what our router sends to Argon.

Argon costs more: $2 per million input tokens and $10 per million output tokens on OpenRouter, against $0.75 and $3.75 for Gemini 3.8 Flash. Both are introductory prices. Google's pricing page says its Argon price rises to $4 and $20 on December 1, 2026, and Google's 3.8 Flash price rises to $1.50 and $7.50 on January 1, 2027. Messages that Gemma answers run on Offline Base's own boards instead of a per-token cloud API.

Why we set reasoning effort to high

OpenRouter's Gemini 4 migration notes say "Gemini 4 Argon always reasons." You choose the effort level (low, medium or high). Both OpenRouter's model listing and Google's guide make medium the default.

Google's guide to Argon recommends medium for most tasks. It describes high as "Most tokens, slowest first token" and suggests it for "Complex reasoning, hard math and science, and the most difficult code or agent tasks." Our router already sends Argon mostly that kind of question, so we set high.

The extra thinking is billed as output. Google's pricing lists output "including thinking tokens," and OpenRouter charges Argon's reasoning at the same $10 per million as its answers. Artificial Analysis counted 110 million tokens from Gemini 4 Argon (High) on its Intelligence Index, against a median of 85 million for comparable models.

What reaches Argon

In Auto mode, EmbeddingGemma 2 on one of our Jetson boards scores each message with the chance that Gemma answers it well. At 50% or higher, Gemma answers. Below that, Argon does. How Gemma Gemini decides who answers each message covers the router in detail. The table shows how the live router scored one example of each kind of message:

Kind of message	Example (Gemma's chance)	Who answers
Proofs and multi-step math	"Prove that the square root of 2 is irrational" (20%)	Argon
Code	Finding a race condition in a Go worker pool (10%)	Argon
Health, legal and money questions	Taking ibuprofen with blood pressure medication (23%)	Argon
Images and PDFs	Any message with a file	Argon, without asking the router
Greetings and chat	"hi there" (80%)	Gemma
Translation and short rewrites	"Translate 'thank you very much' to Japanese" (90%)	Gemma
Current facts	"What's the weather in NYC today?" (94%)	Gemma, after a web search

The scores follow the labels we trained the router on. We also labeled long analysis, research across many sources and complaints about an earlier answer for Gemini. Short questions can go either way: we labeled "What is 15% of 80?" for Gemma and "Is P equal to NP?" for Gemini.

Some messages skip the router. This Gemma setup reads text only, so an image or PDF goes to Argon even when the person picked Gemma. Picking Gemini 4 Argon in the model picker, or "Ask Gemini instead" on the latest reply, sends a message straight there. And when a health check finds no online board with Gemma, messages go to Argon until a board is back.

The router makes mistakes in both directions. In testing, EmbeddingGemma 2 sent some short rewrite requests to the Gemini side that Gemma could have handled, and it kept a chest-pain question on Gemma at 51%. We replaced our cloud router with EmbeddingGemma 2 has the full results.

How long Argon takes to answer

Gemma on a Jetson Orin Nano writes about 25 to 29 tokens per second, and with the model warm its first word usually arrives within 0.2 to 1.3 seconds. Argon thinks first. In our testing, a full proof took about 14 seconds end to end, and Google's guide warns that hard prompts "can take several minutes." The app gives Argon up to two minutes to start answering.

That wait is worth it for a proof or a question about mixing medications. One setting covers every Argon turn, so a plain greeting also runs at high effort when it lands there because the person picked Argon or the boards were down. Small models can handle most everyday questions makes the case for keeping the fast path as the default.

Failover in both directions

The server holds each model's stream until the first word of its answer or its thinking. If the chosen model fails before then, the other one answers as if it had been picked first, and the reply says so: "Gemma 4 E2B was picked, but its host didn't answer, so Gemini 4 Argon answered."

From Gemma to Argon, failover covers a board that is unplugged, busy or stuck. Gemma gets 45 seconds to start and no retry. The app then marks Gemma as down, and once 8 seconds have passed the next message checks the fleet again.

From Argon to Gemma, it covers a cloud call that can't connect, has its key rejected, runs out of credit or gets rate-limited. Argon gets one retry. Gemma is skipped as the backup when a file sent the message to Argon or when Gemma is already known to be down.

OpenRouter adds a layer of its own, serving Argon from two Google providers with automatic failover between them. Once either model has sent its first word, the turn stays with that model.

See it in the route trace

On gemmagemini.com, ask for a proof and open the chip under the reply. The trace shows Gemma's chance against the 50% bar, the time to first word and the model id google/gemini-4-argon. Argon's thinking for that answer is under "Show thinking". Then choose "Ask Gemma instead" to see how the small model handles the same question.