Gemini 4 Argon or Carbon for coding? Cost and tokens from our first tests

TL;DR: in our tests one page generated by Gemini 4 Argon cost between $0.096 and $0.133, and one by Gemini 4 Carbon between $0.122 and $0.176. On the three shared tasks Carbon wrote 34% more tokens, so at the same price it cost 34% more. These are first results from a run that is still going, not a ranking.

You can open every finished page from both models, with its cost and token count, at testingmodels.com/gemini-4.

How we measured

• One completed sample per task (n=1). No tools, no reruns for a better result and no fixes to the code. 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 these initial completed-page costs. The counts and costs are an initial snapshot of completed samples from November 12, 2026.
• The same briefs as every other model. Testingmodels.com is run by Patryk Raba, the publisher of vibecoding.inc, so these are our own tests.
• Gemini API on the Flex tier. Thinking level MEDIUM and a 65,536-token output cap that includes thinking. Introductory price until November 30, 2026: $1 per million input tokens and $5 per million output tokens, then $2 and $10.
• No generation time. Flex requests can wait in a queue, so wall-clock time would mix waiting with the model's work.

Three shared tasks

• Driftwood landing page: Argon 26.6k tokens for $0.133, Carbon 35.1k for $0.176
• NEON BREAKER game: Argon 21.5k for $0.108, Carbon 24.4k for $0.122
• Coffee-chain analytics dashboard: Argon 22.3k for $0.112, Carbon 34.9k for $0.175

Carbon reports thinking tokens separately and they are included. For Argon the API returned no such count, so its figure is the billed output. That is a reporting difference, not proof that Argon did no thinking.

On screen the two took the same brief in different directions. On the landing page Argon writes about photos forgotten in the cloud and shows a photo book cover from Oaxaca, while Carbon opens with the camera roll as a graveyard of great trips and a Kyoto spread. For the dashboard Argon chose a dark theme with a list of cities, Carbon a light one with a donut chart.

Argon on nine tasks

Argon also finished the color palette tool, the particle sandbox, the WebGL 3D racer, the CHIP-8 emulator, the news front page and the digging game. Across nine tasks it wrote 210,979 tokens for $1.06 in total. Every page rendered in our browser check without console errors, but that is a runtime check, not a review of every feature.

What we do not know yet

• Two runs ended with a broken response stream: Carbon on the palette tool and Argon on the kanban board. We record them as infrastructure events, not model results.
• Carbon has only 3 of 40 tasks so far, and just three briefs are shared.
• Gemini 4 Hydrogen, Mistral Large 4 and GPT-6.1 Sol have passed the landing-page browser check and are now live in the arena. Sol used medium effort through the Codex CLI subscription. Claude Opus 5.5 has no finished result yet: it now runs under a separately labeled method that allows up to two automatic continuations of the same prompt in the Claude Code CLI.

If you are choosing a model for simple pages and prototypes, both cost a fraction of a dollar per page on Flex. Before you decide, open both outputs of the same brief side by side.