Two weeks ago, Gemini 4 Carbon was a model most people could read about and nobody outside a vetted defender program could call. That changed on Tuesday. Google opened the model it calls its strongest to every paying API developer and enterprise, and to subscribers on its top consumer tier, thirteen days after the gated debut.

Read the access list for what it excludes. The free API tier is out; Google's pricing page says so in a callout. Consumer access is announced for AI Ultra subscribers, and 9to5Google reports that AI Pro subscribers, who received Argon in October, stay on Argon. "Available to developers, enterprises, scientists, and consumers" is accurate, and narrower than it sounds. The scientists in that sentence reach Carbon through an application program, free for six months if accepted, not through the storefront.

The discount expires on November 30

Carbon launches at Argon's introductory price, $2 per million input tokens and $10 per million output tokens. The introductory part has a date on it. Google's pricing page lists $2.00 through November 30, 2026 and $4.00 starting December 1 for input, $10.00 and then $20.00 for output, on the Standard tier. Argon's card carries the identical schedule.

The doubling is only half the cost story. Google's own post recommends Argon for the majority of software engineering and knowledge work, and describes Carbon as the model that spends more tokens per job, exploring and verifying at length. Same sticker price, larger token appetite, by design. Anyone evaluating Carbon this month should compare representative workloads, including retries and review time, at both the introductory and the scheduled December Standard rates before drawing conclusions.

The scoreboard moved the day before

At its October debut, Carbon's 59 on the Artificial Analysis Intelligence Index put it first, one point clear. It opens to the public one day after Anthropic's Fable 5.2 reset that board: VentureBeat's November 10 read of the index has Fable at 65 on maximum effort, Carbon at 60 on high, Opus 5.5 at 59 on maximum. Carbon is the strongest model Google sells. It is no longer the top score on the chart it launched with.

Google's post leads with three other first places: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, each "at more than 2X lower cost per task compared to the next best external model." The post gives no score and no cost figure for any of the three. VentureBeat reported that the public CursorBench table it reviewed carried no Carbon row and the WANDR results a blank Carbon entry, so those two claims currently rest on Google's own materials. VentureBeat reports published Zapier figures supporting a Carbon performance lead; that does not by itself verify the cost advantage attached to the claim.

The fourth benchmark in the post is new. Google partnered with Artificial Analysis to launch the Multimodal Knowledge Agent benchmark alongside the model, and reports Carbon second on it, with "more than 6X lower average cost per task" than the top model. A benchmark a vendor helped launch can still be useful. Until its methodology and full results are public and reproducible, its first rankings are an opening position, not a settled table.

What a sensible trial looks like

The genuinely new tool in this launch is `/swarm`, in preview in Antigravity: one prompt becomes a coordinated team of agents, with Carbon assigned to architecture, coordination and auditing while faster, lighter models do parallel implementation work in isolated workspaces, checked by adversarial verifiers. Google's own layout puts the token-hungry model on the supervisory roles and cheaper models on the volume work, which is a reasonable way to run it and a hint about how Google expects the bills to look.

So the trial writes itself. Keep Argon, or your current model, as the baseline. Give both systems the migrations, debugging sessions and document work that already consume senior time. Count failures, retries and reviewer hours alongside tokens, and run the arithmetic at $4 and $20, not at the launch discount. The introductory rate ends on a printed date, November 30. What Carbon costs after that is also printed, and it is the number that matters.

Sources

  1. [1] Google, "Gemini 4 Carbon is now available" (November 10, 2026) — primary source; GA announcement, access surfaces, introductory price, Argon recommendation, `/swarm` preview, benchmark claims, MKA partnership and the scientist program; read in full first-handRead source
  2. [2] Google, Gemini API pricing page — Carbon and Argon Standard-tier rates with the November 30 / December 1 schedule and the free-tier exclusion; read first-hand on November 10, 2026Read source
  3. [3] VentureBeat, "Google Gemini 4 Carbon is now publicly available — and it undercuts Claude Fable 5.2 and Opus 5.5 on API price" (November 10, 2026) — Artificial Analysis index positions and the public-table checks for the CursorBench, WANDR and Zapier claimsRead source
  4. [4] 9to5Google, "Gemini 4 Carbon rolling out to Gemini app for AI Ultra, Argon remains" (November 10, 2026) — rollout surfaces; AI Pro subscribers stay on ArgonRead source
  5. [5] ALFW, "Gemini 4 Carbon Leads the AI Index by One Point. Access Starts With Cyber Defenders." (October 28, 2026) — prior ALFW coverage of the announcement and the defenders-only phaseRead source