# Gemini 4 Carbon makes Argon a harder sell **Summary:** Gemini 4 Carbon scores nine points above Argon on Artificial Analysis v5 at nearly identical benchmark task cost. I would test it as an Argon replacement before moving work from other models. - Canonical: https://markhuang.ai/news/gemini-4-carbon-argon-upgrade - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-10-29 - Section: News - Tags: Gemini 4 Carbon, Gemini 4 Argon, Artificial Analysis, Model Evaluation, AI Pricing - Source: [Artificial Analysis](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![A black carbon crystal and a violet-lit glass ampoule sit on the level pans of a brass balance scale](https://cdn.markhuang.ai/news/gemini-4-carbon-argon-upgrade/hero.webp) *Conceptual illustration: Carbon and Argon share token rates. The buying question is what those rates get you.* [Artificial Analysis puts Gemini 4 Carbon first on Intelligence Index v5](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1), with a score of 59 at high reasoning effort. That is nine points above Gemini 4 Argon. The comparison I find more useful sits beside the ranking: Carbon costs $7.76 per index task at standard pricing, against Argon's $7.68. Once I can get access, I'd give Carbon first consideration for Argon's hardest jobs. Higher measured capability at nearly unchanged benchmark spend is a good reason to revisit an incumbent. The case for keeping Argon would be practical: faster completion, perhaps, or better results on the particular work. Both would need testing. This matters most to a team already considering Argon. Carbon gives it an upgrade candidate without a higher token rate to explain. For a team happy with another model, the case depends much more on which tasks it needs done. ## The extra capability comes at Argon's rates Artificial Analysis lists the same standard prices for Carbon and Argon: $4 per million input tokens and $20 per million output tokens. Equal token rates alone would not convince me. A model can use more tokens to finish the job. Here, the reported cost per index task is also nearly equal, which makes the capability gain worth examining. Carbon leads four of the index's nine evaluations, with strengths in complex reasoning, knowledge, and scientific tasks. The [v5 announcement](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5) assigns equal 25% weights to knowledge work, coding, science, and general capabilities. That blend is useful for a broad shortlist. My workload could have a very different mix. I'd keep the version number attached to the comparison, too. v5 changes the evaluation suite. The nine-point gain here compares Carbon and Argon within v5; comparing Carbon with an Argon score from an earlier index would measure a different thing. ## The other incumbents get their own comparison The same Carbon report puts it second on the Coding Agent Index, behind Claude Opus 5.5. That comparison pairs Carbon with Antigravity CLI and Opus with Claude Code. It describes the tested agent setups as well as the models. Before replacing a coding setup, I would want evidence from that setup rather than the overall intelligence ranking. The office-work result is less flattering. Carbon scores 1609 Elo on AA-Briefcase, an improvement over Argon's 1483, but still below models including Kimi K3.1 Pro at 1678. Someone buying an agent mainly for that work has a reason to look beyond the overall winner. The price comparison changes with the alternative. Carbon's $7.76 per index task is below Opus 5.5 at maximum effort, at $10.04, but above GPT-6 Astra at maximum effort, at $6.25. These are benchmark task costs. The cost of a finished report or an accepted patch in my application remains a separate question. Against Argon, then, Carbon offers a higher overall score for similar measured spend. Against other providers, I'd need to establish that its strengths match the job closely enough to justify changing a working setup. ## The introductory rate cannot set the budget [Google's launch announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) prices Carbon initially at $2 per million input tokens and $10 per million output tokens. Its footnote gives the later rates as $4 and $20. Artificial Analysis uses the standard rates throughout its comparison and says the 50% launch discount has no confirmed end date. My budget would use the standard rates. The discount could make an initial evaluation cheaper, but I would want the upgrade to remain worth paying for when it ends. Access is the more immediate constraint. Google says Carbon initially goes to a subset of cyber defenders, followed by additional trusted partners through its Fairwind Program. Broader access is still forthcoming. In the [Hacker News discussion of the ranking](https://news.ycombinator.com/item?id=50288019), several commenters question the usefulness of celebrating a model they cannot try. I share the practical concern: I can plan a comparison now, but my own results will have to wait for access. [TestingCatalog's prelaunch coverage](https://www.testingcatalog.com/gemini-4-carbon-early-test-results/) also cautions that early impressions came from arena matchups and hand-picked prompts, with uncertainty about the checkpoint served. I'd leave latency as an open question until I can measure the service available to me. ## What would earn Argon a place? Once access permits, I would compare the two on the same difficult tasks, with the same acceptance criteria. Failed attempts would go on the bill, along with the time someone spends checking the output. If Carbon produces more acceptable work for similar cost and delay, I would move those tasks. Argon could keep work where it proves faster or cheaper per accepted result. Carbon's overall win leaves room for that. And my [earlier look at Carbon's reported cheating attempts](https://markhuang.ai/news/gemini-4-carbon-cheating-attempts) still applies: better scores would not remove my independent checks. I find this a more consequential result for Argon's place in a workflow than for any claim about one best model. For an Argon workflow, I'd start with the hardest jobs and see whether Carbon makes them easier to finish. Keeping the old model would then need a reason I could measure.