Availability and positioning
Google has made Gemini 4 Carbon generally available at the end of a roughly two-week testing phase reserved for selected security experts. According to The Decoder, the model does not replace Gemini 4 Argon, which was released earlier: Google identifies Argon as the preferred choice for most software engineering and knowledge-work tasks, while Carbon is intended for tasks that push the model's intelligence to its limits.
The distinction matters because the two models have the same per-token price, but Carbon scores higher on benchmarks. Google cites complex multi-step coding and debugging and open-ended scientific research among Carbon's strengths. However, the model consumes more tokens than Argon to reason more deeply and for longer verification cycles: the total cost per task may therefore be higher at the same unit price.
Cost comparison and benchmarks
In the Artificial Analysis test cited by the source, Carbon costs $7.76 per solved task, calculated at the standard price of $4 per million input tokens and $20 per million output tokens. In the same test, Opus 5.5 Max comes in at $10.04, GPT-6 Astra Max at $6.25, and Fable 5.2 at up to $26.50 in the top tier, around $9 in the High tier. In that index, Fable 5.2 nonetheless beats Carbon by two points, 62 to 60.
Google claims new results: Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, with a cost per task less than half that of the runner-up, Fable 5.2. In the new Multimodal Knowledge Agent benchmark, developed with Artificial Analysis for documents dense with charts, diagrams and tables, Carbon ranks second behind Fable 5.2 and ahead of Opus 5.5, but with an average cost less than one-sixth that of the first-placed model. Access is available to developers and business customers via API, AI Studio, Android Studio and Gemini for Business, with an introductory price of $2 per million input tokens and $10 per million output tokens.
Agents and security
In parallel, Google is previewing the /swarm feature in Antigravity, its alternative to Claude Code: a single prompt is transformed into a coordinated team of agents whose work can be tracked in real time. Carbon takes on roles requiring deep reasoning, such as architectural design, coordination and auditing, while lighter models operate in parallel in isolated environments and as adversarial verifiers. The Decoder notes that the cost-benefit ratio of these agent swarms is debated in the industry, with cautious positions from OpenAI and Anthropic and skepticism among researchers; the main benefit appears to be speed through parallelization, but managing them is complex.
On the security front, Google reports results from the early-access phase: as part of Wiz's Scan for Good program, Carbon reportedly identified real zero-day vulnerabilities while analyzing two large SaaS providers in under two hours, while Google's cloud security team reportedly used it to write working proofs of concept for two known bugs in an open-source project. Google also says the model reduces detection errors by 25% or more compared with Argon in the Project Fortify system. Specific cybersecurity benchmark data are not shared and apply only to the version available through the Fairwind program. Its effectiveness outside these controlled contexts remains to be verified.
