Clarity for AI buildersSponsored by AI Routing
availability

Google makes Gemini 4 Carbon generally available, with deeper reasoning and higher token use

Google makes Gemini 4 Carbon generally available, with deeper reasoning and higher token use
Google’s Gemini 4 Carbon is generally available, complementing Argon rather than replacing it, The Decoder reports. The release follows a two-week testing phase with selected security experts. According to The Decoder’s account of Google’s announcement, paid-tier developers and business customers can access Carbon through the Gemini API in Google AI Studio, Android Studio, or Gemini for Business. Google positions the model for demanding reasoning tasks, including complex, multi-step coding, debugging and open-ended scientific research. That positioning comes with an important qualification: Google still recommends Gemini 4 Argon for most software engineering and knowledge work. Carbon and Argon have the same per-token prices, but Carbon uses more tokens for deeper reasoning and longer verification loops. The practical choice is therefore not simply between two models at an identical price. It is whether the additional reasoning improves a particular task enough to justify the larger token bill. Pricing and benchmark trade-offs The Decoder reports an introductory Carbon price of $2 per million input tokens and $10 per million output tokens. Separately, the Artificial Analysis results it cites used standard pricing of $4 per million input tokens and $20 per million output tokens. Those two pricing bases should not be treated as interchangeable when estimating deployment costs. In that Artificial Analysis test, Carbon cost $7.76 per solved task. Fable 5.2 at its High tier cost about $9 per task and scored 62 on the overall index, compared with Carbon’s 60. At its maximum tier, Fable 5.2 cost $26.50 per task. The comparison illustrates why both reasoning settings and task-level spending matter: a higher-scoring configuration can also carry a substantially larger bill. The competitive picture varies by benchmark. According to the report, Google cites Carbon as the leader on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench. On the new Multimodal Knowledge Agent benchmark, developed by Google with Artificial Analysis for dense documents containing charts, blueprints and tables, Carbon finished just behind Fable 5.2 and just ahead of Opus 5.5. Its average cost was less than one-sixth of the top performer’s, the report says. These results support a narrower conclusion than an across-the-board performance lead. Carbon’s relative appeal depends on the workload, the competing model’s configuration and the cost measure used. The supplied report does not include the complete benchmark methods or underlying results needed to independently assess those comparisons. Agent teams remain a preview Alongside Carbon’s availability, Google is previewing a “/swarm” feature in Google Antigravity. The Decoder describes it as turning a single prompt into a coordinated team of agents whose work users can follow in real time. Carbon handles reasoning-heavy roles such as architecture, coordination and auditing, while lighter models work in parallel isolated workspaces and act as adversarial verifiers. Google’s demonstrations include an explorable 3D campus game, a reconstruction of the Eiffel Tower’s construction and a playable virtual fortepiano. These are demonstrations of the preview, not evidence of general availability or measured productivity gains. As the report notes, parallel agents may improve speed, but their cost-benefit balance remains debated. The supplied evidence does not establish whether the extra coordination and verification work reduces total spending. Security results carry scope limits The Decoder also reports Google’s findings from early-access security testing. Through Wiz’s Scan for Good program, Carbon reportedly found zero-day vulnerabilities while scanning two large SaaS providers in under two hours. Google says Carbon improved detection rates in its Project Fortify autonomous security system by more than 25 percent over Argon. Some claimed cybersecurity benchmark leads apply specifically to the version available through the Fairwind Program, rather than necessarily to the generally available model. Google did not provide specific scores for those leads in the supplied report, limiting direct comparisons. For prompt-injection resistance, Google reports a zero percent attack success rate on the Gray Swan IPI benchmark. The result is a claim about that evaluation, not proof that prompt injection has been solved. The Decoder explicitly raises the possibility that the test may be saturated. Users should therefore distinguish reported benchmark resistance from a guarantee that an agent handling untrusted material cannot be compromised.