Gemini 4 Carbon vs Gemini 4 Argon: What Changed in Two Weeks

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon launched in mid-October, and Artificial Analysis scores Carbon 59 on its Intelligence Index v5 versus Argon's 50, ranking Carbon first of 52 models. Carbon also cuts indirect prompt-injection attack success on Gray Swan's test from 0.7% to 0.0% and raises AA-Omniscience accuracy from 48% to 58%, but its hallucination rate rises from 4% to 11%, and Vals AI reports Carbon uses 2.3 times Argon's reasoning tokens and 64% more output tokens on long jobs. Carbon is initially limited to selected cyber defenders with wider rollout through Fairwind, at the same $2/$10 introductory and $4/$20 standard price per 1M tokens as Argon, and no technical report was found for Carbon.

Gemini 4 Argon launched in mid-October 2026. Two weeks later, on October 28, Google DeepMind announced Gemini 4 Carbon https://explainx.ai/blog/gemini-4-carbon-launch-cyber-defenders-first-59-intelligence-index-pricing-2026 and called it its strongest model yet. If you had just wired Argon into a product, the natural question is blunt: did you already fall behind, and is it worth waiting? The short answer is that Carbon is clearly better on the Artificial Analysis Intelligence Index 59 against 50 and on prompt-injection resistance, but it trades some of Argon's honesty for that score, costs more tokens on long jobs, and you probably cannot call it yet. This post lays out the differences row by row, then says what to do about them. TL;DR: Carbon vs Argon | Question | Gemini 4 Argon | Gemini 4 Carbon | |---|---|---| | Announced | Mid-October 2026 | October 28, 2026 | | Who can use it? | Generally available | Selected cyber defenders; wider rollout through Fairwind | | Intelligence Index v5 | 50 high | 59 1st of 52 | | AA-Omniscience accuracy | 48% | 58% | | AA-Omniscience hallucination rate | 4% | 11% | | Indirect prompt-injection attack success | 0.7% | 0.0% | | Price per 1M tokens in / out | $2 / $10 intro, then $4 / $20 | $2 / $10 intro, then $4 / $20 | | Cost per Index task standard price | $7.68 | $7.76 | | Text Arena | 1,520 | 1,528 | | WebDev Code Arena | baseline | about 1,768 +89, 4th | | Technical report | Model card and safety report | None found for Carbon | What got better in Carbon? The gains are real, and several come from outside Google. - Overall capability. Artificial Analysis scores Carbon 59 on Intelligence Index v5, up from Argon's 50, and ranks it first. At its high reasoning tier it holds the top score in four of the nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. - Coding agents. In Google's Antigravity CLI, Carbon's agentic coding score rises from 53 to 59, as reported by The Decoder. That still trails Claude Code with Claude Opus 5.5 by three points, but it is a six-point climb. - Web development. Carbon lands around 1,768 on the WebDev Code Arena, about 89 points above Argon and two behind Claude Sonnet 5.5. - Knowledge work. On AA-Briefcase it gains 126 Elo, from 1,483 to 1,609. - Text quality. Arena has Carbon first in Text Arena at 1,528, eight points above Argon. - Prompt injection. On Gray Swan's indirect prompt-injection test, attack success falls from 0.7% to 0.0%. Google itself says that does not mean prompt injection is solved, and the test looks saturated. - Cheating on Terminal-Bench 4.0. Google's Logan Kilpatrick says successful cheating fell to 0.0% with attempts at 0.4%. That is his statement and we have not seen the underlying data. The Argon launch post https://explainx.ai/blog/gemini-4-argon-launch-benchmarks-pricing-2026 has the baseline for each of these if you want the starting point. What got worse, or at least more expensive? Two things move in the wrong direction, and a fast-moving news cycle tends to bury both. Hallucination rate nearly tripled. Artificial Analysis reports Carbon's AA-Omniscience hallucination rate at 11%, up from Argon's 4%. It is still the second lowest of any model on the list, behind only Argon, and far below the 50% reported for Claude Opus 5.5 and GPT-6 Astra at max effort. But it moved the wrong way. Accuracy rose from 48% to 58%, so Carbon answers more questions and is wrong more often when it does not know. Artificial Analysis found it guesses more when unsure. Accuracy and hallucination balance explaining the Gemini 4 Carbon vs Argon tradeoff Long jobs can burn more tokens. Vals AI reports that Carbon uses 2.3 times Argon's reasoning tokens and 64% more output tokens on its benchmark runs, and that Finance and Tax Agent tasks take more than twice as long. Legal Research averages over 50 minutes per task. Vals puts Carbon's cost at $10.51 per test with a run time just under 48 minutes. Carbon is still below half of Sonnet 5.5's and Opus 5.5's output volume, but the per-token price being unchanged hides a bigger bill on some workloads. On the Artificial Analysis Index itself the two are nearly level on cost: $7.76 per task for Carbon and $7.68 for Argon. So the worry is specific to long, tool-heavy jobs rather than a blanket price rise. Same price, different access Carbon and Argon share an introductory price of $2 per million input tokens and $10 per million output tokens , with cached input 95% off, rising to $4 and $20 at standard pricing. Google has not published when the introductory rate ends. Artificial Analysis uses the standard price in its comparisons, so budget at $4 and $20. Access is where they differ. Argon is generally available. Carbon is not. Google started with a subset of cyber defenders and will add trusted partners through the Fairwind Program, then open it to developers, businesses and consumers after more feedback and guardrail work, with no date. Application runs through a Google Cloud account team and prioritises governments and national cyber authorities, critical infrastructure operators and core technology platforms. Applicants face background checks, phishing-resistant MFA and user-level authentication, and use is limited to internal security, incident response or penetration testing. Access cannot be shared or resold. Responsible scaling ladder with a gate showing why Gemini 4 Carbon access is staged Carbon is also going through the U.S. government's voluntary pre-release testing process. If you have no Google Cloud account team and no security use case, plan as if you will not have Carbon for a while. Is Carbon a new model or a tuned Argon? Nobody outside Google knows. The company published no technical report or parameter count for Carbon, and no separate Carbon system card was found as of October 29. Ars Technica notes Carbon could be a post-training upgrade of Argon rather than a new architecture. The two-week gap fits that reading, but it is an inference, not a fact. This matters for safety. The closest published risk document is Argon's safety report https://explainx.ai/blog/gemini-4-argon-safety-report-activation-probes-misalignment-monitor-2-percent-cheating-2026 , which estimates that without mitigations an attacker could complete a severe historical cyberattack at 16.7 times lower cost. Google has not published an equivalent figure for Carbon, and its headline cyber scores SEC-bench Pro 88.2%, CWE-Bench 71.0%, ExploitGym 54.8% come from Google's own runs without safety guardrails. Read them as claims. Should you build on Argon, wait for Carbon, or use something else? Here is the decision as we see it, from a builder's chair. 1. You are shipping a product now. Build on Argon or a live frontier model such as Claude Opus 5.5 https://explainx.ai/blog/claude-opus-5-5-launch-benchmarks-pricing-2026 . Keep the model name in a config value so a Carbon swap is a one-line change. 2. You do retrieval, support, or anything where a wrong answer is costly. Argon's 4% hallucination rate is a feature. Carbon's 11% is still low by the field's standards, but you should re-run your own evals before swapping. 3. You are a security team. Apply to Fairwind. Carbon's strength is vulnerability discovery, patch writing and long-horizon agent work, and that is the use case Google gated it for. Run your own red-team prompts instead of trusting the 88.2% headline. 4. You run long agent loops. Pin max output tokens , log reasoning-token spend per task, and test Carbon on your hardest ten jobs before you move volume. The cost differences show up there, not on a one-line prompt. 5. You run coding agents. Neither Gemini 4 model leads yet. Claude Code with Opus 5.5 scores 62 on the Coding Agent Index against Carbon's 59. For cross-lab numbers, see our Carbon vs Opus 5.5 vs Astra vs Grok 4.7 comparison https://explainx.ai/blog/gemini-4-carbon-vs-claude-opus-5-5-vs-gpt-6-astra-vs-grok-4-7-comparison-2026 and the earlier Argon four-way https://explainx.ai/blog/gemini-4-argon-vs-opus-5-5-vs-grok-4-7-vs-gpt-6-astra-2026 . What people are asking Did Google release Carbon because Argon disappointed? There is evidence of internal skepticism about Argon's coding performance in a Bloomberg report we covered in Google's Bloomberg skepticism post https://explainx.ai/blog/google-gemini-4-bloomberg-internal-coding-skepticism-2026 , but nothing links that to Carbon's timing. Google calls Carbon its strongest model and says it is state of the art on cybersecurity and long-horizon capabilities. Is the 59 vs 50 gap a fair comparison? Mostly. Both scores come from Artificial Analysis's v5 index at the high reasoning tier. Index v5 was rebuilt on October 27 with a new methodology, so do not compare either number with scores from older index versions, such as the ones quoted in Argon's launch coverage. Will Argon get cheaper or be retired? Google has made no statement. Treat Argon as a stable production target until it says otherwise. Where do I follow access news? Google's Fairwind Program page https://deepmind.google/fairwind-program/?ref=explainx is the official source. As of October 29 it still described Argon access and did not mention Carbon, so check the announcement post and your account team too. Related reading Scores, prices and access are accurate as of October 31, 2026 and come from Google's announcement, Artificial Analysis, Vals AI and Arena as reported that week. They may change as Carbon access widens.