Lançamento de modeloImportante

Google opens Gemini 4 Carbon to paying developers, at an introductory price that doubles after November 30

Two weeks after a defenders-only debut, Google's heaviest reasoning model reaches the Gemini API, Gemini for business and AI Ultra. Google says Argon should remain the default for most work.

Promptea Editorial4 min de leitura

Google made Gemini 4 Carbon broadly available on Tuesday, two weeks after announcing it and limiting first access to vetted cybersecurity defenders through its Fairwind Program. As of today, any paying Gemini API developer can call it from Google AI Studio and Android Studio, enterprises get it through Gemini for business, and Google AI Ultra subscribers can use it in the Gemini app, AI Mode in Search and the Antigravity agent platform, according to Google's announcement.

Carbon is not a replacement for Gemini 4 Argon, the frontier model Google shipped last month. Google's own positioning is explicit: Argon stays the default for most software engineering and knowledge work, while Carbon is meant for the hardest jobs (long multi-step coding and debugging, open-ended scientific research) and deliberately spends more tokens than Argon exploring, reasoning and running verification loops.

Price: cheap per token, for three weeks

Carbon launches at an introductory $2 per million input tokens and $10 per million output tokens, the same introductory rate Google used for Argon. A footnote in the announcement sets the end date: after November 30, 2026, the price becomes $4 input and $20 output per million tokens. The Verge reports that even the post-promo rate undercuts competing frontier models per token, but adds the caveat that matters: a model that thinks longer can cost more per task even at a lower per-token price.

That caveat is built into Carbon's design. The model card says the model uses more thinking tokens than Argon, offers configurable thinking levels to trade quality against latency and cost, and warns that it may sometimes be slow, time out, or use more tokens to maximize performance, especially at higher thinking levels.

What Google claims, and what is not yet verified

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than twice lower cost per task than the next best external model. It also introduced a new Multimodal Knowledge Agent (MKA) benchmark, built with Artificial Analysis, on which it says Carbon ranks second while costing more than six times less per task than the leader. These are vendor-reported figures on benchmarks chosen by the vendor, one of them co-designed for this launch. Independent replication is not yet available, and The Verge notes that companies routinely pick results that flatter their models.

Google's strongest cybersecurity claims (leading results on vulnerability finding, exploitation and patching benchmarks) apply specifically to Carbon accessed through Fairwind, the restricted program for verified defenders, not necessarily to the public API model.

The technical envelope

  • Context: up to 1M tokens, counted as combined input plus output, per the model card. Long reasoning traces and long outputs eat into the same budget as your prompt.
  • Modalities: text, image, audio and video in; text out.
  • Knowledge cutoff: March 2026.
  • Lineage: the model card describes Carbon as based on Gemini 4 Argon and defers to Argon's card for architecture, training data and acceptable use.
  • Agent tooling: Google is previewing /swarm in Antigravity, which turns a single prompt into a coordinated team of agents, with Carbon intended for roles such as architecture, coordination and auditing and lighter models doing parallel work.

Safety numbers, all self-reported

Google reports that in simulated cyberattack environments Carbon's safeguards blocked 99% of attack trajectories from an adaptive red-teaming agent and cut the per-attempt breach rate to 0.8%, from 3.1% on Argon. It cites external testing by Gray Swan (automated cyberattack query success of 2.17% versus 9.86% for Argon) and Trajectory Labs (no new universal jailbreaks in 132 hours and more than 13,000 requests), plus an internal agentic-honesty benchmark where Carbon passed 92% of 375 tasks against 70% for Argon. Google also repeats its earlier claim of a 0% attack success rate on Gray Swan's indirect prompt injection benchmark, while acknowledging that its own red teaming shows the model can still be improved against sophisticated attacks.

What changes for builders

The practical decision is routing, not replacement. Google itself is telling developers to keep Argon as the workhorse and reserve Carbon for tasks where extra reasoning pays off. If you adopt Carbon, measure cost per completed task rather than per token, set the thinking level deliberately instead of defaulting to the maximum, and leave headroom in the 1M window for output and reasoning, since both count against it.

Plan for the price change too. Any cost model built on the introductory rate will be off by a factor of two from December 1. Prototyping during the promo window is cheap; committing production budgets to it is not a safe assumption.

Access is still tiered. API use requires the paid tier, consumer access requires Google AI Ultra, and The Verge notes Google has not said when, or whether, lower subscription tiers or free users will get the model. Google's developer post points to an updated Gemini 4 developer guide for migration details.

Por que isso importa

  • A third frontier option now sits on a public API at a low per-token price, which changes model-routing and cost decisions for teams running coding and research agents.
  • Carbon is designed to spend more reasoning tokens, so per-token price alone is a poor guide to what a task will actually cost.
  • The introductory price expires on November 30, 2026, so budgets built on it will double from December 1.

Pontos principais

  • Available from November 10, 2026 on the paid Gemini API tier (AI Studio, Android Studio), Gemini for business and Google AI Ultra.
  • Pricing: $2 input / $10 output per million tokens until November 30, 2026; $4 / $20 afterwards.
  • Model card: up to 1M tokens combined input and output, multimodal input, text output, March 2026 knowledge cutoff.
  • Benchmark leadership and safety improvements are Google's own figures; independent evaluation is not yet available.

Fontes

  1. GooglePrimária
    Gemini 4 Carbon is now available
    blog.google
  2. GooglePrimária
    Gemini 4 Carbon is now available for developers
    blog.google
  3. Google DeepMindPrimária
    Gemini 4 Carbon — Model Card
    storage.googleapis.com
  4. The Verge
    Google releases Gemini 4 Carbon with agent 'swarm' feature
    theverge.com

Como escolhemos e verificamos as fontes

Tags:
  • gemini
  • model selection
  • api pricing
  • context windows
  • coding agents
  • cybersecurity
Empresas:
  • Google
  • Google DeepMind
Modelos:
  • Gemini 4 Carbon
  • Gemini 4 Argon
Gemini 4 Carbon is now available: pricing, context and caveats · Promptea