Google Opens Gemini 4 Carbon at $2 per Million Input Tokens, a Quarter of Anthropic's Fable 5.2 Price

Google made Gemini 4 Carbon available to paying developers on November 10. The general release comes two weeks after the October 28 announcement, when Google limited access to selected cyber defenders and partners in its Fairwind program. According to MLQ, access now covers the Gemini API, enterprise customers, Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Google's Antigravity coding environment. Pro and free consumer tiers are excluded.

Pricing matters most to companies that build products on top of the model. Cursor's model documentation lists Google's introductory rate of $2 per million input tokens and $10 per million output tokens through November 30. From December 1 the rate rises to $4 input, $0.20 cached read and $20 output. One day earlier, Anthropic cut Claude Fable 5.2 to $8 input and $40 output. Carbon therefore costs one quarter of Fable per token at the introductory rate and one half at the standard rate.

The capability evidence is close and mostly comes from vendors. Cursor reported that Carbon at high effort scored 60.4% on CursorBench, the company's internal coding benchmark, against 60.0% for Fable 5.2 Max. The Russian-language channel AI ML Big Data summarized Google's claim that Carbon ranks first on three agentic benchmarks: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Anthropic's own launch table shows Carbon ahead on AutomationBench at 53.9% to 51.4%. The same table shows Carbon behind on FrontierCode 2.0, at 41.1% against Fable's 49.1%. Artificial Analysis scored Carbon 59 on its Intelligence Index, and Fable 5.2 later scored 65. A 0.4-point lead on one benchmark run by a vendor does not establish a ranking. Cursor also has not published the per-task cost data that would support Google's claim that Carbon costs less than half as much per completed task.

Google also previewed "swarms" in Antigravity, where one prompt launches a team of agents. AlphaCorp warns that swarms use quota much faster than a single chat session. That makes the per-token rate less useful than cost per completed task as a basis for procurement.

Google has not disclosed Carbon's architecture or how it achieves the price. Research posted on November 11 shows one way labs reduce long-context serving cost. The ADIndex paper targets DeepSeek Sparse Attention (DSA). In DSA, the indexer still scores the entire causal prefix for every query, and at long context that indexer becomes a dominant source of latency. The authors propose layer-adaptive indexing trained by group-level distillation. Earlier 2026 work on the same bottleneck reported 2x kernel speedups at 32K and 4x at 128K with hierarchical indexing. IndexCache removed 75% of indexer computation on a 30B DSA model and reported a 1.82x prefill speedup. No source connects these methods to Carbon.