Gemini 4 Carbon: $2/$10 Intro Pricing Until Nov 30, Carbon vs Argon, and /swarm in Antigravity

Google made Gemini 4 Carbon generally available on November 10, 2026. Google calls Carbon its strongest model, a step above Gemini 4 Argon, and it’s now open to all paying developers and enterprises through Google AI Studio and Gemini for business. Two weeks earlier, Google had released it only to cybersecurity defenders in its Fairwind Program.

The price is the part with a deadline. Until November 30, 2026, Gemini 4 Carbon costs the same as Argon: $2 per million input tokens and $10 per million output tokens. From December 1, 2026, Google’s pricing doubles to $4 input and $20 output. If you want to benchmark Carbon on your own coding tasks, the rest of November is the cheap window.

Google also previewed /swarm in Google Antigravity, a feature that turns one prompt into a coordinated team of agents, with Carbon taking the roles that need the deepest reasoning.

Gemini 4 Carbon pricing: intro rate vs December rate#

Google’s “What’s new in Gemini 4 Carbon” page lists these paid-tier rates per million tokens:

Three billing details from the same page change how far that rate goes:

• Thinking tokens bill as output. Carbon spends more thinking tokens per turn than Argon on complex tasks, so a hard prompt can cost more on Carbon even while the per-token price is identical.
• Grounding adds tokens. Google counts internal tool system instructions and function declarations when you use grounding, and Grounding with Google Search is still billed per query.
• No free tier. Gemini 4 Carbon isn’t available on the Gemini API Free tier.

Google’s blog footnote adds that the $4/$20 price applies to the Carbon model intended for public access only.

Carbon or Argon: what Google says each is for#

Google doesn’t tell everyone to move to Carbon. Its launch post says Argon remains the ideal choice for the majority of software engineering and knowledge work. Carbon is for tasks that test the limits of a model’s intelligence, such as complex, multi-step coding and debugging and open-ended scientific research.

The two models share the same 1M-token unified context window, the same built-in tools and the same per-token price. The developer docs put the difference in thinking tokens: Carbon spends more of them per turn to explore and verify its work, which raises both token use and latency compared to Argon.

Google’s docs list these uses for Carbon: complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing. For interactive coding, day-to-day feature work and fast production loops, the same page points you back to gemini-4-argon.

The benchmark claims, with their source#

All of the numbers in this section come from Google’s announcement. They are company claims, not independent results.

• Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than 2x lower cost per task than the next best external model.
• On the new Multimodal Knowledge Agent (MKA) benchmark, built with Artificial Analysis, Carbon ranks #2 at more than 6x lower average cost per task than the top model.
• On Google’s 375-task agentic honesty benchmark, Carbon passed 92% of tasks, against 70% for Argon.
• Google says Carbon reached a 0% attack success rate on the Gray Swan indirect prompt injection (IPI) benchmark. In the same section, Google says prompt injection robustness remains an area for continued improvement.

Google’s comparison is cost per task, not cost per token. Carbon uses more tokens per turn than Argon, so the per-token price alone won’t tell you what a task costs. Check it on your own workload before you budget around it.

Calling gemini-4-carbon from the API#

Switching an Argon app to Carbon needs no request schema changes, according to Google’s migration notes. Change the model string from gemini-4-argon to gemini-4-carbon. Google’s quickstart uses the Interactions API:

curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "model": "gemini-4-carbon",
    "input": "Analyze this payment processing pipeline for race conditions during retry attempts and rewrite the transaction locks safely.",
    "generation_config": {
      "thinking_level": "medium"
    }
  }'

thinking_level is the main cost and latency control. Google recommends medium (the default) for most tasks, including complex code and agentic use, low for simpler and cost-sensitive work, and high for the hardest reasoning. Google’s advice is to run your evaluations at all three levels and pick the lowest one that meets your quality bar.

Coming from Gemini 3 or earlier, there’s more to change. Google’s migration checklist includes:

max_output_tokens caps thinking and output together. If you set it too low, Carbon can spend the budget on thinking and cut the answer short. Google’s suggested starting points are the default 65,536 for standard tasks, and 131,072 or 262,144 with background execution for complex reasoning and coding.

Implicit caching is on by default for prompts of at least 4,096 tokens, and Google says it discounts cached input by 95%. Agent loops that resend a large, stable prefix (system prompt, repo map, tool definitions) are where that discount shows up.

/swarm in Google Antigravity#

Google is previewing /swarm in Google Antigravity. Google says it turns a single prompt into a coordinated team of agents you can track in real time. Developers can assign Carbon to roles that need the deepest reasoning, such as architectural design, coordination and auditing, and pair it with faster, lighter models. Those agents work in parallel isolated workspaces, with adversarial verifiers that test every change.

Google’s own examples are large builds from one prompt: an interactive Three.js reconstruction of an office from seven public planning PDFs over 10 overnight loops, and a 3D browser game built by coordinating subagents across Blender, Three.js and Web Audio. Treat these as Google’s demos, not as typical results.

/swarm is a preview. In Antigravity and the Gemini app, Carbon is rolling out to Google AI Ultra subscribers. For API users, the Antigravity agent in Gemini Managed Agents also supports gemini-4-carbon through agent_config.

The architect-plus-cheaper-workers pattern isn’t unique to Google. Anthropic describes a similar orchestrator setup for Claude Fable 5.2, which we covered in Claude Fable 5.2 and Mythos 5.2 pricing.

Where Gemini 4 Carbon is available#

• Developers: the Gemini API paid tier in Google AI Studio and Android Studio. Google AI Ultra subscribers can also try it in Google Antigravity and AI Studio.
• Enterprises: Gemini for business via the API or model drop-down, Google Antigravity, or CodeMender.
• Consumers: the Gemini app and AI Mode in Search for Google AI Ultra subscribers.
• Scientists: Google’s free-access program for researchers now includes Carbon alongside Argon, subject to eligibility, capacity and selection.
• Security teams: the Fairwind Program gives access to Carbon for cyber on a rolling basis.

Is Gemini 4 Carbon worth testing now?#

If you have hard, multi-step coding or debugging tasks where Argon or another model stalls, test Carbon before December 1, 2026. Through November 30 you pay Argon’s $2/$10 rate, so the comparison costs less now than it will next month. Measure cost per finished task, not per token, because Carbon’s extra thinking tokens bill as output.

If Argon already handles your work, Google itself says to keep it for most software engineering. Carbon’s per-turn token use and latency are higher, and from December the per-token price doubles too. For a wider look at what each provider charges, see our cheapest AI coding API comparison.