Gemini 4 Carbon:#1 on the index, behind Opus 5.5 on code.

The real problem (beyond the announcement)

Two weeks after Gemini 4 Argon, Google DeepMind announced Gemini 4 Carbon on 28 October 2026, positioning it as its strongest model with an explicit focus on cybersecurity and long-horizon tasks. For an engineer, the real question is not “is it the best model in the world”. It is: is it the best one for my workload, at what cost, and can I call it today?

Short answer: not yet. Carbon is not open to developers. Google is giving it first to a subset of cyber defenders through its Fairwind Program, as part of the U.S. government's voluntary pre-release testing process. Developer access is promised “as soon as possible”, with no date. If your roadmap depends on a model you cannot call, that is not a roadmap, it is a waiting list.

Second point: topping an aggregate index does not mean topping your workload. Carbon scores 59 on Artificial Analysis Intelligence Index v5, updated the same week. As Nathan Lambert details on Interconnects, the index weights four blocks equally — knowledge work (25%), coding (25%), science (25%), general (25%) — and any aggregation into a single number makes trade-offs. Your task distribution has no reason to be balanced the same way.

What the benchmarks say (and what they cost)

Google's cybersecurity numbers are clean: 88.2% on SEC-bench Pro (finding real vulnerabilities), 54.8% on ExploitGym (validation through exploitation), 71.0% on CWE-bench (root-cause patching) — all state of the art. Internally, Google reports a 2x uplift on secure code generation over Argon, and the Chrome team reports patches of “10x better” quality than another large frontier model. Those benchmarks were evaluated without guardrails, which is both more honest and more uncomfortable.

On general coding, the picture is more nuanced — and the announcement spends less time there. From the independent evaluations compiled by Latent Space:

In other words: first on aggregate indexes and on cybersecurity, third behind Opus 5.5 and GPT-6 Astra on real code work, and behind several models — including one open-weight, Kimi K3.1 Pro — on knowledge-work tasks.

Pricing is the strongest argument. Carbon will launch at Argon's introductory price: $2 per million input tokens, $10 per million output tokens, with cached input at 95% off ($0.20/M). After the introductory period: $4 / $20. In practice, Artificial Analysis measures $7.76 per index task (vs $10.04 for Opus 5.5), with 29% fewer output tokens than Opus. But Vals AI observes the opposite on finance, tax and legal tasks: 64% more output tokens than Argon. Cost per token does not give cost per task: measure it on your own workloads.

Two honest signals not to skip. First, the hallucination rate measured by Artificial Analysis rose from 4% (Argon) to 11% (Carbon) on AA-Omniscience — still among the lowest on the market (Opus 5.5 and Astra sit at 50%), but nearly triple the previous model in the same family. Second, output limits: Carbon accepts 1M tokens of context and up to 1M output tokens, with text, image, video and speech input — but it only produces text.

On agentic behavior and internal safety, the numbers Google published deserve to be read for what they are — vendor claims, but quantified. Carbon ranks first on Vending-Bench 2 and Blueprint-Bench 2 (Andon Labs), where it matches the human baseline on 3D spatial understanding, first on Text Arena (1528 Elo) and second on Agent Arena ($1.80 median cost per task). Google also reports 0.0% successful cheating on Terminal-Bench 4.0 (0.4% attempts) and about 2x fewer unsupported claims than Argon. Metrics to replay on your own stack before believing them.

When to use it (and when to skip it)

Worth waiting for:

• Long agentic runs: Google reports multi-day runs hunting subtle bugs across large codebases. If your agents give up after two hours, this is the capability being pushed.
• The cybersecurity chain: detection, validation, patching. 88.2% / 54.8% / 71.0% are the most credible numbers in the announcement, on public benchmarks.
• High-volume knowledge work, where cost per task ($7.76 vs $10.04) and cached input (−95%) change the economics.

Skip it for now:

• If your need is pure code: Opus 5.5 and GPT-6 Astra still lead on FrontierSWE and FrontierCode. A “#1 index model” that loses by five points on code is a bad pick for a development team.
• If you need multimodal output: Carbon only produces text.
• If you must ship this week: access is closed. Benchmark Argon (available, same introductory price) and keep your abstraction layer ready.
• If your evaluation rests on a single aggregate score: read AA v5's weighting (25/25/25/25) and replay your own tasks. A model can be first on the index and third on your actual pipeline.

Production checklist

1. Do not migrate to a gated model. Benchmark on Argon today, keep an abstraction (model router, provider SDK) so switching on release day needs no rewrite.
2. Replay your evaluation suite on your task distribution. AA v5's four blocks weigh 25% each; yours do not. Compare Carbon and your current model on 30 to 50 real examples, not on a leaderboard.
3. Budget for verbosity. 64% more output tokens than Argon on legal and finance tasks (Vals AI): set max_output_tokens, measure cost per task, not per million tokens.
4. Keep a factuality guardrail. 11% hallucination on AA-Omniscience: retrieval grounding, mandatory citations on factual answers, human review on decisions.
5. For a cybersecurity pipeline, keep humans on patches. 71% on CWE-bench means three patches in ten fail — and the announced benchmarks were run without guardrails.

Need help designing reliable AI workflows, auditing an LLM architecture, or training your engineering team? Explore how we work at Kaleido-Lab.