Gemini 4 Carbon Is an Announcement, Not Yet a Switching Decision

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it its next frontier model and its strongest model. The announcement lists an introductory price of $2 per million input tokens and $10 per million output tokens. Access is staged, and the results cited here are claims from Google’s announcement, not buyer-specific evaluation results. ¹ That is enough to open vendor diligence. It is not enough to approve a production switch.

The distinction matters for anyone who owns a model budget. An announcement can trigger an evaluation. It cannot tell you whether a model will beat your current one on your workloads, meet your operational requirements, or cost less per accepted result. Google has published benchmark scores, internal testing claims, a stated price, and an access plan. It has left the questions that matter to a buyer open. ¹

What did Google DeepMind announce with Gemini 4 Carbon?

DeepMind calls Carbon its next frontier model and its strongest model — the company’s own characterization. It also says Carbon ranks first with a score of 59 on the v5.0 Artificial Analysis Intelligence Index. ¹

The announcement reports results in cybersecurity, web development, coding safety, and unsupported claims. These are areas Google says Carbon improved; they do not establish how it will perform across an enterprise’s full workload mix. ¹

On commercial terms, the release gives an introductory price and a staged access plan. The announcement does not specify context limits, rate limits, which modalities are supported, deployment requirements, or a named developer API for accessing Carbon. ¹

Analysis: Treat this as a product signal and a diligence trigger, not evidence that Carbon beats an incumbent model or lowers operating cost.

What do the reported benchmarks show — and what do they not show?

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching, evaluated without guardrails according to the company. ¹ Teams doing security-related work may find these results relevant. They do not show how Carbon behaves inside a buyer’s own tools, permissions, and review process. For a comparable read on vendor security claims, see GPT-6.1 Sol in CodeRabbit's Security Agent: What Buyers Need to Verify.

On WebDev Arena, Google reports Carbon gained 89 Elo points and ranks fourth at 1768. ¹ That is a result from one evaluation setting, not a substitute for testing the applications a company needs to support.

Internally, Google says Carbon achieved more than a 2x uplift in secure-code-generation success rate over Argon. Its Chrome team reportedly found Carbon more than 10X better than an unnamed large frontier model from another lab at generating high-quality patches. ¹

The announcement does not name that comparison model or provide the evaluation methods, datasets, or settings needed to reproduce the claim. Google also reports that Carbon cuts unsupported claims by approximately 2x over Argon in internal testing. ¹ These are vendor-reported figures, not a procurement benchmark that buyers can reproduce from the information provided.

Analysis: Keep vendor-reported numbers separate from results your own team reproduces. A strong score in one test tells you what to evaluate next. It does not replace the evaluation.

Does the Gemini 4 Carbon price make it cheaper to use?

Google says Carbon’s introductory price matches Argon’s: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. ¹

That is a starting point for a cost model, not an answer to what completing a business task will cost. You would still need your own input-to-output ratio, cache behavior, traffic pattern, latency requirement, and quality bar to turn the rates into a meaningful estimate.

Google describes these as introductory rates but does not say how long they will hold. ¹ Procurement should record that uncertainty and seek written commercial terms before committing budget.

Analysis: Compare cost per accepted result, not token price alone. A model that needs more retries, writes longer outputs, or requires extra verification can cost more at a lower rate. The announcement provides no buyer-specific evidence on those factors. ¹

Can an enterprise team use Carbon now?

Google says it would start making Carbon available to a subset of cyber defenders during the week of the announcement, then gradually extend access to trusted partners through its Fairwind Program. Broader access for developers, businesses, and consumers through Google’s product surfaces is planned after further feedback and guardrail work. ¹

The announcement gives no calendar schedule beyond that rollout sequence, eligibility criteria, or specific access mechanism for a buyer. It also does not name a developer API. ¹ An announcement by itself does not tell a given company whether it can get access or run a production-like trial.

Google also says it is participating in the U.S. government’s voluntary pre-release testing process, without sharing results or specifying what evidence buyers will eventually see. ¹

Analysis: Make access a gate in the evaluation plan. Before committing engineering time, confirm how your organization can request access, what environment it runs in, and whether it supports the tests you need.

What evidence is still missing before an enterprise model switch?

Reproducing or interpreting the reported results requires details such as datasets, prompts, scoring rules, model configurations, sampling settings, and baselines. Those details are not provided in the announcement excerpt. ¹

Google also describes internal use: its teams working with Carbon for days on complex, multi-step engineering tasks, including root-cause analysis and patching subtle bugs in large codebases and low-level systems. ¹ That is a report of internal use, not an independent evaluation or a forecast of performance in another organization’s environment.

Operational details such as context limits, rate limits, supported modalities, access mechanisms, latency, and deployment restrictions are not specified in the announcement excerpt. ¹ Request these details before scoping a trial.

Security claims need the same treatment. Google says Carbon builds on Argon’s safety and security stack, including model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. ¹ A description of a safety stack is not a test against your threat model.

Analysis: Track evidence in separate buckets:

• Google-reported: results and product descriptions from the announcement.
• Independently reproduced: results your own team can rerun under documented conditions.
• Buyer workload: performance on your own tasks, data, tools, and review criteria.

Keeping these apart stops a headline score from quietly becoming a production assumption. The same discipline applies to other vendor launches; see Gemini's Overnight ADK Optimizations: What Companies Should Verify.

How should platform teams run an enterprise model evaluation of Gemini 4 Carbon?

Start with a decision gate tied to a real workload: run Carbon against the incumbent and a current alternative using identical prompts, data conditions, tool permissions, output limits, and review criteria. This is a proposed method, not a claim about how Carbon will perform.

Measure task success and failure recovery alongside factual accuracy, security defects, patch acceptance, latency, token consumption, and cost per accepted result. Google reports fewer unsupported claims from Carbon than Argon in internal testing; each team should check whether that difference appears on its own workload. ¹

For coding and security use, test both defensive and adversarial cases. Google reports results in vulnerability finding, exploitation, patching, cheating behavior, and internal misalignment alerts. ¹ Those results do not establish how Carbon behaves with your specific tools, permissions, and data.

A staged trial can limit the risk of mistaking a strong demo for production readiness:

• Offline replay: Run a fixed set of representative tasks against each candidate.
• Controlled review: Score outputs against shared acceptance criteria.
• Limited shadowing: Observe performance on a narrow slice of live traffic without letting Carbon make decisions.
• Scoped rollout: Expand only once predefined quality, security, and cost thresholds are met.

One practical aid: if candidates are reachable through a single enterprise AI gateway and listed in a governed model catalog, the same prompts can be sent to each with one API integration, which makes side-by-side comparison easier to keep consistent. Whether Carbon is reachable that way depends on how and when Google makes it available.

Analysis: Set exit criteria before the trial starts. If Carbon does not improve an outcome the business cares about — or introduces unacceptable security or operational cost — the team should be able to stop the trial rather than let it drift into a default migration.

What should procurement and platform leaders do next?

Ask Google to fill the gaps: evaluation protocols, the identity and configuration of the unnamed comparison model, access criteria and timing, API details, operational limits, and how long the introductory prices will hold. ¹

Build a cost-and-quality trial using the published rates — $2 per million input tokens, $10 per million output tokens, and 95% off the input token price for cached input — but label them introductory in every model. ¹ Write down assumptions about input/output mix and what counts as a successful result before running the numbers.

Do not approve a migration on the strength of a title, benchmark rank, or price card. Approve a bounded evaluation once the team has a defined workload, access to the model, a repeatable baseline, and an exit criterion tied to business impact.

Next step: assign an owner in platform engineering and an owner in procurement to request access and commercial terms, then put a workload-specific evaluation plan in front of the production decision-maker.

Sources

¹ “Gemini 4 Carbon: Our next frontier model,” Google DeepMind, published October 28, 2026. https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/