Gemini 4 Argon and Carbon: First One-Shot Results, Cost and Tokens

Google made Gemini 4 Argon generally available on October 13, 2026 and Gemini 4 Carbon on November 10, 2026. Both are now in the arena, running the same published one-shot briefs every other model gets. This is a first report from a run that is still going: Argon has finished 9 challenges, Carbon 3, and every complete output is live and unedited on the Gemini 4 results page.

How these were run

• Route: Gemini API, Flex tier, the discounted queue-based service. Introductory Flex price for both models is $1 per million input tokens and $5 per million output tokens until November 30, 2026, then $2 / $10 (Google AI Studio pricing). Thinking tokens are billed as output.
• Settings: thinkingLevel MEDIUM on every request, maxOutputTokens 65,536 (a hard cap that includes thinking), the arena system prompt and the unchanged task brief. No tools.
• One completed sample per task (n=1): no reruns for a better result and no edits to the HTML. Infrastructure retries are recorded separately: an HTTP 503 on the Argon particle task and an HTTP 504 on the Carbon color task were retried. The incomplete responses are preserved separately, and their unresolved billing is excluded from the initial completed-page costs below. These counts and costs are an initial snapshot of the completed Argon and Carbon samples from November 12, 2026.
• Cost: computed from the usage the API returned at the Flex price on the run date. We do not publish generation time for these runs, because Flex requests can wait in a queue first.

Same three challenges, both models

On these three briefs Carbon wrote about 34% more billed output than Argon, and cost about 34% more, because both use the same price. Carbon reports its thinking tokens separately (1,691, 1,096 and 3,546 here), and they are included above. For Argon the API returned no thinking-token count, so its figure is simply the billed output. That is a reporting difference, not proof that Argon did no thinking.

What we see on screen, not a verdict: on Driftwood, Argon pitches "Your trips deserve better than a forgotten cloud folder" with a terracotta photo-book cover, while Carbon goes with "Your camera roll is a graveyard of great trips" and an open Kyoto field-notes spread. On the dashboard brief they made opposite calls: Argon built a dark dashboard with a ranked city list, Carbon a light one with a category donut.

Argon on nine challenges

Argon has completed the landing page, the arcade game, the color tool, the dashboard, the particle sandbox, the WebGL tunnel racer, the CHIP-8 emulator, the news front page and the digging game. Across those nine it wrote 210,979 billed output tokens for $1.06 in total, between 19,206 and 26,600 tokens and $0.096 to $0.133 per page. All nine rendered in our browser check without console or page errors. That is a runtime check, not a review of every feature: play them in the arena before you trust any one of them.

What did not finish

• Two runs ended with a broken response stream: Carbon on the color tool and Argon on the kanban board. They are recorded as unresolved infrastructure outcomes while billing is reconciled. They are preserved separately from completed model results, with each infrastructure attempt recorded.
• Carbon has finished three challenges so far. The rest of its 40-brief suite is queued behind the other new models.

Also new in the arena

GLM-5.3 finished the Driftwood brief through its coding plan, with thinking on: 58,176 output tokens, of which 31,752 were reasoning. That run was on a subscription, so no per-run cost is shown. Compare it with Carbon.

Gemini 4 Hydrogen, Mistral Large 4 and GPT-6.1 Sol have completed their landing pages, passed the browser checks and are live in the arena. Open their Driftwood outputs alongside GLM-5.3. Sol used medium effort through the Codex CLI subscription; its method is labeled separately from the API runs.

The full method is on the method page. New results land on the Gemini 4 page as they pass their checks.