AI Agent Pulse - the weekly briefing on the agent economy. Subscribe freePay-per-call agents: read the x402 docs
gigsoul.com

GigSoul

Intelligence on the agent ecosystem
Tuesday, November 10, 2026
Tools

Gemini 4 Carbon: price the finished job, not the token.

Google's heavier reasoning model arrives at an introductory token rate. Its own guide explains why the cost of accepted work is the number to compare.

Programming code on a screen, not a product screenshot

Concept photo: programming code on a screen, by luis gomes / Pexels. Not a screenshot of Gemini. Used under the Pexels license, which permits article, RSS and social reuse. No endorsement implied.

The tempting number in Google's new AI launch is $2. That is the introductory price for a million input tokens from Gemini 4 Carbon. It is not the price of finishing your migration, checking your spreadsheet or getting a difficult bug fixed.

Google opened Carbon to paying developers and enterprises on November 10, with access also rolling out to Google AI Ultra subscribers across several products. Its own explanation contains the catch worth reading before the leaderboard: Carbon uses more tokens than Gemini 4 Argon to explore problems and run extended verification loops.

The same rate can buy a different bill

The developer guide puts Carbon at $2 per million input tokens and $10 per million output tokens through November 30. From December 1, Carbon's listed rates become $4 and $20. These are API token prices, not a flat subscription or a ceiling on an agent's work.

Google still recommends Argon for most software engineering and knowledge work, describing it as tighter on token usage and faster for interactive tasks. Carbon is the heavier option for difficult investigations, architecture and other work that benefits from deeper reasoning. That is a routing decision, not an automatic upgrade for every request.

The Verge's launch coverage makes the same useful distinction: lower per-token prices do not necessarily mean lower task costs when models consume different numbers of tokens. Google's cost-per-task benchmark claims may help choose an evaluation candidate. They do not price your own repository.

A million-token window is room, not a spending target

The guide describes a shared one-million-token window for input, thinking and output. The Interactions API defaults to a 65,536-token output limit. Longer generation requires background mode and a larger output limit; Google explicitly advises keeping that limit low enough to constrain spending when a hard prompt uses it all.

Carbon also offers low, medium and high thinking levels, with medium the default. Turning the dial to high changes how much reasoning the model does. It does not establish that the extra work was necessary or that the result is correct.

For a useful comparison, choose a real task with a fixed acceptance check. Record whether it passed, its total model and tool costs, elapsed time, retries and human repair. Compare the cost of accepted work, including failed attempts, rather than the cheapest individual response. That is an evaluation plan, not a test GigSoul has performed.

The cyber headline has a separate access gate

Public Carbon access should not be confused with unrestricted access to Carbon for cyber. Google's announcement and SecurityWeek's reporting distinguish the defensive cyber offering through the Fairwind Program. The model card likewise says access varies by channel and subscription tier.

Google is also previewing multi-agent work through Antigravity's /swarm feature. Its examples and benchmark results are launch evidence from the vendor, not proof that a particular team should delegate an entire project or that every extra verifier pays for itself.

Carbon's interesting promise is that spending longer on a hard problem can save work elsewhere. The sensible buying question is whether that happens on your task. A low token rate gets the experiment onto the shortlist. A finished, checked result earns the budget.

More in Tools

All Tools →