Google Gemini 4 Carbon Goes Public at a Quarter of Claude Fable 5.2's API Price

Quick Facts

• Google launched Gemini 4 Carbon to paid API and Google AI Ultra subscribers on Nov. 10, 2026, after a two-week limited release to select cybersecurity firms.
• Introductory pricing starts at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 on Dec. 1.
• Claude Fable 5.2, which launched a day earlier, costs $8 per million input tokens and $40 per million output tokens.

Google opened Gemini 4 Carbon to all paying developers on Nov. 10, 2026, positioning the model as a lower-cost alternative to Anthropic's Claude Fable 5.2 and OpenAI's GPT-6 Astra. The model is available through the Gemini API in Google AI Studio, Gemini for business, and to Google AI Ultra subscribers at $100 per month via the Gemini app, AI Mode in Search, and Google Antigravity.

The Gemini API has no free tier for Carbon. Pro and free Gemini users are excluded.

Pricing Window Is Short

At launch, Carbon costs $2 per million uncached input tokens and $10 per million output tokens. That rate is a quarter of Claude Fable 5.2's $8 input and $40 output pricing, which itself dropped from $10 and $50 for Fable 5.1. The introductory rate also amounts to roughly one-fifth of OpenAI's GPT-6 Astra.

The window closes fast. On Dec. 1, Google doubles both rates to $4 per million input tokens and $20 per million output tokens. Cached input runs about 95% cheaper than standard input at list price.

After the December price increase, Carbon will cost half what Fable 5.2 charges per token, not a quarter. Teams building cost-sensitive pipelines have roughly three weeks to lock in the lower rate.

Built for Cybersecurity and Long-Horizon Engineering

Google describes Carbon as its strongest model yet for cybersecurity defense and long-horizon agentic workflows. According to Google's announcement, the model uses more thinking tokens than its predecessor Gemini 4 Argon, with configurable thinking levels to balance quality, latency, and cost.

The model accepts text, images, audio, and video, with a combined input-output context window of up to 1 million tokens. Internally, Google's teams report Carbon running for days on complex, multi-step engineering tasks including root-causing and patching bugs across large codebases.

On cybersecurity benchmarks, Carbon ranks first on SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation, and CWE-bench at 71.0% for patching. In simulated cyberattacks, the model blocked 99% of attack paths from adaptive red-teaming agents and cut the breakthrough rate per attempt from 3.1% with Argon to 0.8%. Google says Carbon discovered zero-day vulnerabilities in major SaaS providers within two hours in internal testing.

Two weeks before the general launch, Google released Carbon to a set of cybersecurity defenders through its Fairwind Program. A free six-month program for scientists also provides access.

Swarm Agents Preview in Antigravity

Google is previewing a feature called /swarm inside Google Antigravity, its AI programming platform. The /swarm command turns a single prompt into a coordinated team of agents that users can track in real time. Google positions this as Carbon's ability to manage large, complex engineering projects across multiple simultaneous workstreams.

Benchmarks Tell a Mixed Story

Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 at a score of 59, and leads on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench. On the Coding Agent Index, Carbon scores 59 using Antigravity CLI, against 62 for Claude Code with Opus 5.5 at maximum settings.

In head-to-head benchmark comparisons, Fable 5.2 leads Carbon in six of seven shared categories. The cost difference runs in the opposite direction. Carbon costs $7.76 per index task, versus $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra.

For software and security teams weighing model costs against capability, Carbon's December price increase makes the Nov. 30 cutoff a real decision point. The question is whether the benchmark gap with Fable 5.2 matters enough to justify paying more once introductory pricing expires.