Gemini 4 Carbon Pricing: Argon's Token Rate, a Bigger Bill
Agentic AIGemini 4 Carbon costs exactly what Gemini 4 Argon costs per token, yet Artificial Analysis put it at $7.76 per solved task, because Carbon spends more tokens thinking. For teams routing agent work between model tiers, that makes the per-token price list the wrong place to compare. Google’s own advice adds a twist: for most software engineering work, it still recommends the cheaper-in-practice Argon.
What Happened
The Decoder reports that Google has made Gemini 4 Carbon generally available after a two-week testing phase limited to selected security experts. Paying developers and business customers can reach it through the Gemini API in Google AI Studio, in Android Studio, or through Gemini for Business. The introductory price stays at $2 per million input tokens and $10 per million output tokens.
Google says Carbon complements Argon rather than replacing it. Argon remains the better choice “for most software engineering and knowledge work tasks,” in The Decoder’s summary of Google’s guidance, while Carbon targets complex multi-step coding and debugging and open-ended scientific research. Google also published a model card and a full safety report.
Why It Matters
The cost gap comes from behavior, not price. Carbon uses more tokens than Argon to reason through hard debugging and run longer verification loops, so The Decoder expects the total bill to be higher despite the identical rate. In the Artificial Analysis test, calculated at the standard price of $4 and $20 per million tokens, Carbon’s $7.76 per task still came in below the older budget model Flash 3.8.
Against rivals, the same test put Opus 5.5 Max at $10.04, GPT-6 Astra Max at $6.25 and Fable 5.2 at its maximum tier at $26.50. At the High tier, Fable 5.2 costs about $9 per task and leads Carbon on the overall index, 62 to 60. Google’s new results claim first place on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at less than half the per-task cost of Fable 5.2. On the new Multimodal Knowledge Agent benchmark, Carbon placed second behind Fable 5.2 at under one-sixth of its average cost.
The safety numbers are the other headline. Carbon passed 92% of a 375-task honesty test, against 70% for Argon, and recorded a zero percent prompt injection success rate on the Gray Swan IPI benchmark. The Decoder cautions that this may mean the test is saturated, not that prompt injection is solved.
What This Means for Builders
Google is also previewing “/swarm” in Antigravity, which turns one prompt into a coordinated agent team. Carbon takes the deep-reasoning roles of architecture, coordination and auditing, while lighter models work in parallel isolated workspaces and act as adversarial verifiers. That is the orchestrator-and-worker split, with the expensive model reserved for judgment. The Decoder notes that OpenAI, Anthropic and researchers have all sounded cautious about whether swarms justify their token use.
Stack Archive’s look at Gemini 4 Argon’s output token habits showed the same lesson from the other side: token volume decides cost as much as rate. The Fable 5.2 per-task analysis makes the same case for Anthropic’s lineup.
The So What
Agent platform teams should benchmark Carbon on their own hardest tasks, such as multi-step debugging and long-document analysis, and compare cost per completed task against Argon and Fable 5.2 High rather than the price sheet. Keep Argon as the default for routine coding, as Google itself advises, and budget for the introductory $2/$10 rate rising to the $4/$20 standard price used in the benchmark.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →